Humaneness Voice Large · Intense bucket adapters

Humaneness Voice Large 1.5 and 1.6 Alpha — model and manual navigation

The merged parent used in these specialist comparisons is now released as Humaneness Voice Large 1.6 Alpha: 1.5 + final RealSpeech rank 4, merged once. Apply only the selected new bucket adapter. Do not reload either ancestor.

1.5 Original model · 1.6 merged model · Complete context manuals for humans and agents

Read 1.5 context manual · Read 1.6 + adapter context manual. Exact prompts, CFG, adapter scales and failed gates are retained. New model/documentation contributions: CC BY 4.0; existing upstream and gated specialist terms remain applicable.

Manual prompts × CFG × bucket adapters

72/72completed; 9adapters currently qualify for an additional held-out benefit. Pending targets have no new recommendation.

The original grid used generic GENERAL/SCRIPT directions at CFG1. It did not systematically use target-specific manual winners. The new study freezes the actual rank-4 manual wording, preserves every literal word, seed, reference, adapter and frame budget, then independently varies CFG and adapter strength. The previous recordings and conclusions remain below.

Ranking:100% requested EARS emotion/style head, standardized using frozen training statistics; low-pole targets reverse the sign. Formal/casual contrasts average the desired head and negative opposite head. CE and genuineness are diagnostics. Higher target z means stronger desired delivery, not a human rating. More CFG is not automatically better.

Selection:20 original-text outcomes per configuration select wording, CFG and scale. Fifty additional unseen-text reference pairs test the frozen choices. A recommended adapter must improve against both the same-CFG no-adapter control and the best manual-only setting. Quality gates: per-take WER≤15%, mean WER≤10%, valid/terminated fraction≥90%, paired identity change≥−0.05 over50pairs. Failed outcomes stay visible. Language-cluster bootstrap intervals use only five fixed reference voices; findings across72targets remain exploratory.

Five languages: EN/DE/FR/NL/ES; original seeds1701/1702; reference187frames/14.96seconds. CFG1/1.25/1.5/2/2.5/3/4/5 × bucket strength0/0.5/0.75/1. CFG4/5 are stress conditions, not presumed improvements. No-reference use of these bucket recommendations has not been validated.

Taxonomy:S_MONO means Monologue, S_TECH means Teacher/Didactic, VOLT means Volatility. The two corrected style recipes and17dimension/contrast recipes are new searches, not historical manual winners.53other targets reuse directly applicable manual recipes.

Agent configuration registry · Detailed agent manual

TargetEvidenceRecommended LoRA strengthCFGManual vs standard ΔzLoRA vs same-CFG ΔzLoRA vs best manual ΔzDecision
Intense Affectiondirect_manual_recipe0.01.00.9030.141-0.064No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Amusementdirect_manual_recipe0.01.01.057-0.0470.294No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Angerdirect_manual_recipe0.752.51.0630.3950.395Held-out target improves against matched CFG and best manual-only configuration; QA passes
Intense Astonishment Surprisedirect_manual_recipe0.03.02.5620.0170.017Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Awedirect_manual_recipe0.04.01.4290.0470.275Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Bitternessdirect_manual_recipe0.03.00.2600.0310.031Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Confusiondirect_manual_recipe0.01.02.3170.1220.122No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Contemplationdirect_manual_recipe0.05.01.322-0.056-0.056Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Contemptdirect_manual_recipe0.01.00.0300.027-0.018No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Contentmentdirect_manual_recipe0.04.00.7460.0110.020Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Disappointmentdirect_manual_recipe0.05.00.424-0.052-0.112Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Disgustdirect_manual_recipe0.04.00.6160.1730.820Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Distressdirect_manual_recipe0.03.02.2750.0641.361Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Elationdirect_manual_recipe0.01.00.299-0.1550.975Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Fatigue Exhaustiondirect_manual_recipe0.04.03.716-0.1890.132Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Feardirect_manual_recipe0.02.50.9720.2752.047Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Helplessnessdirect_manual_recipe0.05.01.7140.0020.002Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Hope Enthusiasm Optimismdirect_manual_recipe0.01.00.436-0.1920.035No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Impatience and Irritabilitydirect_manual_recipe0.01.01.9100.2570.989No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Interestdirect_manual_recipe0.02.50.5640.030-0.131Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Longingdirect_manual_recipe0.05.01.421-0.062-0.062Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Paindirect_manual_recipe0.754.00.4940.3630.741Held-out target improves against matched CFG and best manual-only configuration; QA passes
Intense Reliefdirect_manual_recipe0.05.00.508-0.011-0.177Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Sadnessdirect_manual_recipe0.05.01.048-0.090-0.090Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Sournessdirect_manual_recipe0.01.250.0040.0760.351Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Teasingdirect_manual_recipe0.02.52.474-0.147-0.147Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Thankfulness Gratitudedirect_manual_recipe0.03.00.0750.0360.036Use manual-only configuration; no clear qualifying incremental adapter gain
Strong ASMR Styledirect_manual_recipe0.01.01.3610.0130.026No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong Authoritative Styledirect_manual_recipe0.05.00.7530.0790.079Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Casual Styledirect_manual_recipe0.04.00.8070.0450.045Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Conversational Styledirect_manual_recipe0.01.00.5270.061-0.102No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong Dramatic Styledirect_manual_recipe0.01.01.3060.0210.021No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong Formal Styledirect_manual_recipe0.04.00.3090.027-0.051Use manual-only configuration; no clear qualifying incremental adapter gain
Strong introspective monologue delivery, not monotone speechcorrected_taxonomy_new_search1.05.00.5540.0950.095Held-out target improves against matched CFG and best manual-only configuration; QA passes
Strong Narrator Styledirect_manual_recipe0.01.00.2880.0250.025No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong Playful Styledirect_manual_recipe0.01.00.3450.0040.510Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Storytelling Styledirect_manual_recipe0.05.00.459-0.009-0.009Use manual-only configuration; no clear qualifying incremental adapter gain
emotionally vulnerable and exposednew_canonical_dimension_extension0.755.01.9110.1610.161Held-out target improves against matched CFG and best manual-only configuration; QA passes
emotionally detached and unexposednew_canonical_dimension_extension0.01.00.1510.0410.041Use manual-only configuration; no clear qualifying incremental adapter gain
dominant and assertivenew_canonical_dimension_extension0.04.00.7460.0590.272Use manual-only configuration; no clear qualifying incremental adapter gain
submissive and deferentialnew_canonical_dimension_extension0.05.01.1650.0700.070Use manual-only configuration; no clear qualifying incremental adapter gain
highly aroused and activatednew_canonical_dimension_extension0.01.250.5410.1140.301Use manual-only configuration; no clear qualifying incremental adapter gain
very low arousal, calm or sleepynew_canonical_dimension_extension0.05.00.669-0.015-0.015Use manual-only configuration; no clear qualifying incremental adapter gain
Emotionally volatile delivery with rapid or abrupt changesnew_canonical_dimension_extension0.03.00.5870.1100.372Use manual-only configuration; no clear qualifying incremental adapter gain
Emotionally steady, stable delivery without abrupt changesnew_canonical_dimension_extension1.05.00.2250.1060.106Held-out target improves against matched CFG and best manual-only configuration; QA passes
strongly positive vocal affectnew_canonical_dimension_extension0.02.50.589-0.0200.109Use manual-only configuration; no clear qualifying incremental adapter gain
strongly negative vocal affectnew_canonical_dimension_extension1.05.00.9980.3080.308Held-out target improves against matched CFG and best manual-only configuration; QA passes
very warm vocal deliverynew_canonical_dimension_extension0.05.00.337-0.002-0.002Use manual-only configuration; no clear qualifying incremental adapter gain
very tense and strainednew_canonical_dimension_extension0.753.00.5910.1700.170Held-out target improves against matched CFG and best manual-only configuration; QA passes
very fast speaking temponew_canonical_dimension_extension0.01.00.1890.0040.066No new configuration passes the held-out gates; retain standard fallback and inspect audio
Very informal/casual, not formalnew_canonical_dimension_extension0.05.00.611-0.090-0.157Use manual-only configuration; no clear qualifying incremental adapter gain
Very formal, not casualnew_canonical_dimension_extension0.01.00.8000.0480.048No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Concentrationdirect_manual_recipe0.01.00.145-0.006-0.084No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Doubtdirect_manual_recipe0.01.01.1570.1390.139No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Embarrassmentdirect_manual_recipe0.04.00.094-0.028-0.042Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Emotional Numbnessdirect_manual_recipe0.55.00.4460.0470.047Held-out target improves against matched CFG and best manual-only configuration; QA passes
Intense Infatuationdirect_manual_recipe0.02.50.259-0.073-0.027Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Intoxication Altered States of Consciousnessdirect_manual_recipe0.04.00.0660.0140.021Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Jealousy and Envydirect_manual_recipe0.04.00.021-0.002-0.002Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Malevolence Malicedirect_manual_recipe0.01.00.153-0.063-0.104No new configuration passes the held-out gates; retain standard fallback and inspect audio
Intense Pleasure Ecstasydirect_manual_recipe0.02.00.8890.050-0.215Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Pridedirect_manual_recipe0.02.00.5640.0870.427Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Sexual Lustdirect_manual_recipe0.02.50.308-0.0160.034Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Shamedirect_manual_recipe0.04.00.038-0.010-0.009Use manual-only configuration; no clear qualifying incremental adapter gain
Intense Triumphdirect_manual_recipe0.01.01.7110.091-1.196No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong Cartoonish Styledirect_manual_recipe0.01.250.3660.0230.023Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Newsreader Styledirect_manual_recipe0.04.00.893-0.0320.174Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Ranting/Angry Styledirect_manual_recipe0.01.02.5600.086-0.335No new configuration passes the held-out gates; retain standard fallback and inspect audio
Strong teacher-like, didactic deliverycorrected_taxonomy_new_search0.02.50.109-0.0340.030Use manual-only configuration; no clear qualifying incremental adapter gain
Strong Whisper-Talk Styledirect_manual_recipe0.01.01.3690.0160.081No new configuration passes the held-out gates; retain standard fallback and inspect audio
very relaxed and unstrainednew_canonical_dimension_extension1.05.00.1550.0370.103Held-out target improves against matched CFG and best manual-only configuration; QA passes
very slow speaking temponew_canonical_dimension_extension0.03.00.1740.0070.007Use manual-only configuration; no clear qualifying incremental adapter gain

This is supervised fine-tuning, not a new reinforcement-learning run. The common baseline is Humaneness Voice Large 1.5 Alpha with the real-speech rank-4 adapter already merged once. Each new bucket adapter is rank32. Scale 0.5/0.75/1.0 multiplies that new adapter’s residual; it is not CFG. CFG is fixed at 1.

Every comparison keeps literal text, English GENERAL/SCRIPT instructions, seed, reference and frame budget identical. Both random seeds are shown; no best-seed cherry-picking. There are two new three-sentence texts per language and five languages. Each language uses one fixed, already-public synthetic reference, 10–15 seconds long. Languages with no accepted bucket-training data are explicitly cross-language transfer tests.

The displayed target score is a normalized EARS head, not an emotion probability. CE/genuineness are model predictions, not human MOS. Means ± sample standard deviations are descriptive; four recordings per language/condition are not a large independent listener study.

Which adapters actually raise the requested target score?

A paired comparison of all 72 trained rank-32 adapters at residual strengths 0.5, 0.75 and 1.0, against the same merged base with the new adapter disabled. This analysis reuses all 5,760 existing recordings; no new generations or best-seed selection were added.

14Clearer gains across languages
40Positive mean, but uncertain
18No positive mean at any strength

The short answer: 54 adapters raise the primary raw target mean at one or more strengths. Only 14 meet the stricter cross-language pattern below. The other 18 do not beat their own baseline on that primary measure at any tested strength. Using the existing CDF-transformed 0–1 mean instead,55 have some increase;52 improve both raw and transformed means at the same strength. The table shows both views and flags disagreements.

Among the 14 clearer adapter results, the selected starting strengths are 0.5 for 1,0.75 for 9 and 1.0 for 4 adapters. Higher strength is not reliably better; the preferred setting depends on the adapter.

Most promising starting points

The following have both a clearer cross-language pattern and a mean gain of at least 0.1 z. This 0.1 cutoff is a display heuristic, not an established threshold for audible improvement. Smaller consistent effects remain visible in the full table.

Astonishment SurpriseStrength 0.75 · +0.375 z5/5 language means improveStrong Authoritative StyleStrength 1.0 · +0.306 z5/5 language means improveStrong ASMR StyleStrength 0.75 · +0.285 z5/5 language means improveStrong Conversational StyleStrength 0.5 · +0.246 z4/5 language means improveStrong Casual StyleStrength 0.75 · +0.239 z5/5 language means improveAffectionStrength 0.75 · +0.203 z4/5 language means improveEmotionally steady, stable delivery without abrupt changesStrength 1.0 · +0.166 z5/5 language means improveEmotionally volatile delivery with rapid or abrupt changesStrength 0.75 · +0.140 z5/5 language means improve

Useful counterexamples: Anger, Amusement, Pride and dramatic style show no average target increase at any of these strengths. A trained adapter is not automatically a better steering option. Even clearer effects can be tiny: Disgust at 0.75 is about+0.016 z and Gratitude at 1.0 about+0.024 z. These score changes are not demonstrated listener-perceived improvements.

How to read the table

  1. Positive Δ z means more of the requested trait. We compare the exact same text, prompt, seed, reference and frame budget. For low-pole requests such as slower tempo or lower tension, the raw head should go down; we reverse its sign so a useful change is still positive.
  2. z is a scale unit, not a probability. It is the raw target change divided by that EARS head’s frozen training standard deviation. We also show the existing 0–1 target score before and after. That score uses a Gaussian-CDF transform and can saturate.
  3. Clearer is stronger evidence within this small grid, not a universal guarantee. The stricter exploratory language-cluster resampling range must stay above zero, and at least 4/5 language means must improve. For the two style-contrast requests, the opposite head must move in the intended direction too.
  4. Strength is not CFG. It multiplies only the new rank-32 adapter residual. The shared 1.5 Alpha base already contains the real-speech rank-4 adapter; CFG is fixed at 1.
  5. Read the quality costs separately. Positive CE/genuineness changes are higher predicted scores; positive WER changes are worse. No quality-weighted reward was used to choose the target-score winner.

Showing 72 of 72 adapters.

Clearer cross-language pattern Mean rises; uncertainty remains No mean increase at this strength

All tested strengths, paired against the same no-adapter baseline. Expand any cell for language results and uncertainty.
Adapter / requested traitStarting strengthCosts at starting strength
Astonishment Surprise
Emotion · emo_Astonishment_Surprise
Raw target direction: ↑		+0.083 z
Mean up; uncertain
Target 0.518 → 0.547
Δ score +0.029
Uncertainty and paired results

Stricter resampling range: [-0.154, +0.381] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.3455 → 0.3912. Desired-direction raw change: +0.0458. CE +0.02; genuineness -0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.115 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.197 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.063 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.417 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.372 z; 50% paired wins; Δ WER +0.00 pp
+0.375 z
Clearer gain
Target 0.518 → 0.612
Δ score +0.095
Uncertainty and paired results

Stricter resampling range: [+0.230, +0.482] z. 5/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 0.3455 → 0.5528. Desired-direction raw change: +0.2073. CE +0.00; genuineness -0.06; WER +0.29 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.409 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.355 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.146 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.481 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.486 z; 100% paired wins; Δ WER +1.47 pp
+0.094 z
Mean up; uncertain
Target 0.518 → 0.561
Δ score +0.043
Uncertainty and paired results

Stricter resampling range: [-0.090, +0.290] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.3455 → 0.3975. Desired-direction raw change: +0.0520. CE -0.04; genuineness -0.05; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.168 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.031 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.390 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.195 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.085 z; 75% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE +0.00
Genuine -0.06
WER +0.29 pp
5/5 language means improve
Strong Authoritative Style
Talking style · vn_S_AUTH_reg
Raw target direction: ↑		+0.201 z
Mean up; uncertain
Target 0.894 → 0.915
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [-0.031, +0.433] z. 4/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 2.6046 → 2.8138. Desired-direction raw change: +0.2092. CE +0.01; genuineness -0.06; WER +0.53 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.157 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.320 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.057 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.529 z; 75% paired wins; Δ WER +2.63 pp
  • ES: +0.258 z; 75% paired wins; Δ WER +0.00 pp
+0.175 z
Mean up; uncertain
Target 0.894 → 0.900
Δ score +0.006
Uncertainty and paired results

Stricter resampling range: [-0.118, +0.476] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.6046 → 2.7859. Desired-direction raw change: +0.1813. CE +0.00; genuineness -0.07; WER +0.79 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.201 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.291 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.271 z; 25% paired wins; Δ WER +3.95 pp
  • NL: +0.629 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.022 z; 50% paired wins; Δ WER +0.00 pp
+0.306 z
Clearer gain
Target 0.894 → 0.931
Δ score +0.037
Uncertainty and paired results

Stricter resampling range: [+0.146, +0.420] z. 5/5 language means improve. 90% of 20 paired recordings improve.

Raw head mean: 2.6046 → 2.9230. Desired-direction raw change: +0.3184. CE +0.01; genuineness -0.10; WER +0.61 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.038 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.379 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.307 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.455 z; 100% paired wins; Δ WER +1.39 pp
  • ES: +0.354 z; 75% paired wins; Δ WER +1.67 pp
Try 1.0
Clearer raw-target gain on this grid		CE +0.01
Genuine -0.10
WER +0.61 pp
5/5 language means improve
Strong ASMR Style
Talking style · vn_S_ASMR_reg
Raw target direction: ↑		+0.189 z
Clearer gain
Target 0.509 → 0.578
Δ score +0.069
Uncertainty and paired results

Stricter resampling range: [+0.057, +0.318] z. 4/5 language means improve. 80% of 20 paired recordings improve.

Raw head mean: 1.6327 → 1.8932. Desired-direction raw change: +0.2605. CE +0.04; genuineness +0.06; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.370 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.007 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.264 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.101 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.218 z; 50% paired wins; Δ WER +0.00 pp
+0.285 z
Clearer gain
Target 0.509 → 0.608
Δ score +0.098
Uncertainty and paired results

Stricter resampling range: [+0.099, +0.461] z. 5/5 language means improve. 80% of 20 paired recordings improve.

Raw head mean: 1.6327 → 2.0253. Desired-direction raw change: +0.3926. CE +0.03; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.407 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.262 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.541 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.009 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.206 z; 75% paired wins; Δ WER +0.00 pp
+0.155 z
Mean up; uncertain
Target 0.509 → 0.563
Δ score +0.054
Uncertainty and paired results

Stricter resampling range: [-0.024, +0.316] z. 4/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.6327 → 1.8461. Desired-direction raw change: +0.2134. CE +0.01; genuineness +0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.334 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.050 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.312 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.113 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.191 z; 25% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE +0.03
Genuine +0.01
WER +0.00 pp
5/5 language means improve
Strong Conversational Style
Talking style · vn_S_CONV_reg
Raw target direction: ↑		+0.246 z
Clearer gain
Target 0.525 → 0.613
Δ score +0.089
Uncertainty and paired results

Stricter resampling range: [+0.017, +0.532] z. 4/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 2.6242 → 2.9453. Desired-direction raw change: +0.3211. CE -0.02; genuineness +0.08; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.051 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.063 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.364 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.678 z; 100% paired wins; Δ WER +0.00 pp
  • ES: +0.177 z; 50% paired wins; Δ WER +0.00 pp
+0.130 z
Mean up; uncertain
Target 0.525 → 0.575
Δ score +0.051
Uncertainty and paired results

Stricter resampling range: [-0.094, +0.471] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.6242 → 2.7941. Desired-direction raw change: +0.1700. CE +0.00; genuineness +0.02; WER +0.87 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.022 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.122 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.161 z; 75% paired wins; Δ WER +1.32 pp
  • NL: +0.677 z; 75% paired wins; Δ WER +1.39 pp
  • ES: -0.086 z; 50% paired wins; Δ WER +1.67 pp
-0.005 z
Mean down / no gain
Target 0.525 → 0.524
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.218, +0.220] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.6242 → 2.6175. Desired-direction raw change: -0.0067. CE -0.06; genuineness -0.05; WER +0.53 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.130 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.060 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.252 z; 75% paired wins; Δ WER +2.63 pp
  • NL: +0.212 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.300 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Clearer raw-target gain on this grid		CE -0.02
Genuine +0.08
WER +0.00 pp
4/5 language means improve
Strong Casual Style
Talking style · vn_S_CASU_reg
Raw target direction: ↑		+0.097 z
Mean up; uncertain
Target 0.180 → 0.218
Δ score +0.038
Uncertainty and paired results

Stricter resampling range: [-0.059, +0.234] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.4067 → 2.5303. Desired-direction raw change: +0.1237. CE -0.04; genuineness +0.01; WER -0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.283 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.145 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.031 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.201 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.113 z; 75% paired wins; Δ WER -1.67 pp
+0.239 z
Clearer gain
Target 0.180 → 0.245
Δ score +0.064
Uncertainty and paired results

Stricter resampling range: [+0.108, +0.324] z. 5/5 language means improve. 75% of 20 paired recordings improve.

Raw head mean: 2.4067 → 2.7119. Desired-direction raw change: +0.3052. CE -0.00; genuineness +0.04; WER +0.19 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.336 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.010 z; 75% paired wins; Δ WER +2.63 pp
  • FR: +0.277 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.318 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.255 z; 75% paired wins; Δ WER -1.67 pp
-0.013 z
Mean down / no gain
Target 0.180 → 0.172
Δ score -0.008
Uncertainty and paired results

Stricter resampling range: [-0.174, +0.146] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.4067 → 2.3897. Desired-direction raw change: -0.0169. CE -0.03; genuineness -0.09; WER +0.47 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.005 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.182 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.165 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.211 z; 50% paired wins; Δ WER +4.02 pp
  • ES: +0.124 z; 50% paired wins; Δ WER -1.67 pp
Try 0.75
Clearer raw-target gain on this grid		CE -0.00
Genuine +0.04
WER +0.19 pp
5/5 language means improve
Affection
Emotion · emo_Affection
Raw target direction: ↑		+0.036 z
Mean up; uncertain
Target 0.701 → 0.682
Δ score -0.019

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.178, +0.192] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.7067 → 0.7262. Desired-direction raw change: +0.0195. CE -0.01; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.117 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.126 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.239 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.015 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.287 z; 50% paired wins; Δ WER +0.00 pp
+0.203 z
Clearer gain
Target 0.701 → 0.763
Δ score +0.063
Uncertainty and paired results

Stricter resampling range: [+0.028, +0.374] z. 4/5 language means improve. 80% of 20 paired recordings improve.

Raw head mean: 0.7067 → 0.8170. Desired-direction raw change: +0.1103. CE +0.03; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.031 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.442 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.279 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.060 z; 100% paired wins; Δ WER +0.00 pp
  • ES: +0.264 z; 50% paired wins; Δ WER +0.00 pp
+0.142 z
Mean up; uncertain
Target 0.701 → 0.721
Δ score +0.021
Uncertainty and paired results

Stricter resampling range: [-0.093, +0.370] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.7067 → 0.7836. Desired-direction raw change: +0.0769. CE +0.02; genuineness -0.11; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.465 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.222 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.182 z; 0% paired wins; Δ WER +0.00 pp
  • NL: -0.033 z; 25% paired wins; Δ WER +1.39 pp
  • ES: +0.236 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE +0.03
Genuine -0.07
WER +0.00 pp
4/5 language means improve
Emotionally steady, stable delivery without abrupt changes
Voice dimension · vn_VOLT_reg_low
Raw target direction: ↓		+0.157 z
Clearer gain
Target 0.935 → 0.952
Δ score +0.017
Uncertainty and paired results

Stricter resampling range: [+0.045, +0.306] z. 5/5 language means improve. 75% of 20 paired recordings improve.

Raw head mean: 1.3478 → 1.2114. Desired-direction raw change: +0.1364. CE -0.04; genuineness -0.04; WER -0.01 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.393 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.012 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.176 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.139 z; 75% paired wins; Δ WER -0.07 pp
  • ES: +0.066 z; 75% paired wins; Δ WER +0.00 pp
+0.084 z
Mean up; uncertain
Target 0.935 → 0.942
Δ score +0.007
Uncertainty and paired results

Stricter resampling range: [-0.041, +0.172] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 1.3478 → 1.2749. Desired-direction raw change: +0.0729. CE -0.04; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.197 z; 75% paired wins; Δ WER +1.39 pp
  • DE: +0.129 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.139 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.072 z; 50% paired wins; Δ WER -1.39 pp
  • ES: -0.117 z; 25% paired wins; Δ WER +0.00 pp
+0.166 z
Clearer gain
Target 0.935 → 0.955
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [+0.121, +0.225] z. 5/5 language means improve. 95% of 20 paired recordings improve.

Raw head mean: 1.3478 → 1.2040. Desired-direction raw change: +0.1438. CE -0.05; genuineness -0.08; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.205 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.117 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.131 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.252 z; 100% paired wins; Δ WER -1.39 pp
  • ES: +0.123 z; 100% paired wins; Δ WER +0.00 pp
Try 1.0
Clearer raw-target gain on this grid		CE -0.05
Genuine -0.08
WER -0.28 pp
5/5 language means improve
Emotionally volatile delivery with rapid or abrupt changes
Voice dimension · vn_VOLT_reg_high
Raw target direction: ↑		+0.008 z
Mean up; uncertain
Target 0.093 → 0.095
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.059, +0.077] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 1.5028 → 1.5094. Desired-direction raw change: +0.0066. CE -0.01; genuineness -0.04; WER -0.24 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.043 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.111 z; 75% paired wins; Δ WER +1.56 pp
  • FR: +0.012 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.083 z; 50% paired wins; Δ WER -2.78 pp
  • ES: +0.041 z; 50% paired wins; Δ WER +0.00 pp
+0.140 z
Clearer gain
Target 0.093 → 0.120
Δ score +0.027
Uncertainty and paired results

Stricter resampling range: [+0.063, +0.208] z. 5/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 1.5028 → 1.6243. Desired-direction raw change: +0.1214. CE -0.06; genuineness -0.07; WER -0.56 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.183 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.225 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.185 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.033 z; 50% paired wins; Δ WER -2.78 pp
  • ES: +0.074 z; 75% paired wins; Δ WER +0.00 pp
-0.029 z
Mean down / no gain
Target 0.093 → 0.088
Δ score -0.005
Uncertainty and paired results

Stricter resampling range: [-0.283, +0.160] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 1.5028 → 1.4776. Desired-direction raw change: -0.0253. CE -0.04; genuineness -0.04; WER -0.56 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.216 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.096 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.105 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.408 z; 25% paired wins; Δ WER -2.78 pp
  • ES: +0.057 z; 75% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE -0.06
Genuine -0.07
WER -0.56 pp
5/5 language means improve
Strong Storytelling Style
Talking style · vn_S_STRY_reg
Raw target direction: ↑		+0.067 z
Clearer gain
Target 0.607 → 0.632
Δ score +0.025
Uncertainty and paired results

Stricter resampling range: [+0.003, +0.178] z. 5/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.7140 → 2.8078. Desired-direction raw change: +0.0939. CE +0.01; genuineness +0.05; WER +0.57 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.065 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.003 z; 50% paired wins; Δ WER -1.32 pp
  • FR: +0.014 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.254 z; 75% paired wins; Δ WER +4.17 pp
  • ES: +0.000 z; 50% paired wins; Δ WER +0.00 pp
+0.091 z
Clearer gain
Target 0.607 → 0.639
Δ score +0.032
Uncertainty and paired results

Stricter resampling range: [+0.011, +0.191] z. 4/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 2.7140 → 2.8404. Desired-direction raw change: +0.1265. CE -0.02; genuineness -0.05; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.243 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.127 z; 75% paired wins; Δ WER -1.32 pp
  • FR: +0.050 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.047 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.014 z; 25% paired wins; Δ WER +0.00 pp
+0.078 z
Clearer gain
Target 0.607 → 0.637
Δ score +0.030
Uncertainty and paired results

Stricter resampling range: [+0.031, +0.147] z. 5/5 language means improve. 80% of 20 paired recordings improve.

Raw head mean: 2.7140 → 2.8224. Desired-direction raw change: +0.1084. CE -0.04; genuineness -0.05; WER +0.07 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.020 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.068 z; 75% paired wins; Δ WER -1.32 pp
  • FR: +0.199 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.038 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.063 z; 75% paired wins; Δ WER +1.67 pp
Try 0.75
Clearer raw-target gain on this grid		CE -0.02
Genuine -0.05
WER -0.26 pp
4/5 language means improve
Sourness
Emotion · emo_Sourness
Raw target direction: ↑		+0.030 z
Mean up; uncertain
Target 0.359 → 0.370
Δ score +0.011
Uncertainty and paired results

Stricter resampling range: [-0.000, +0.065] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0573 → 0.0702. Desired-direction raw change: +0.0128. CE +0.01; genuineness -0.02; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.028 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.016 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.027 z; 50% paired wins; Δ WER +1.39 pp
  • NL: +0.020 z; 50% paired wins; Δ WER -0.07 pp
  • ES: +0.089 z; 75% paired wins; Δ WER +0.00 pp
+0.056 z
Clearer gain
Target 0.359 → 0.380
Δ score +0.021
Uncertainty and paired results

Stricter resampling range: [+0.038, +0.075] z. 5/5 language means improve. 75% of 20 paired recordings improve.

Raw head mean: 0.0573 → 0.0816. Desired-direction raw change: +0.0243. CE -0.00; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.037 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.056 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.065 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.038 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.085 z; 100% paired wins; Δ WER +0.00 pp
+0.059 z
Mean up; uncertain
Target 0.359 → 0.381
Δ score +0.022
Uncertainty and paired results

Stricter resampling range: [-0.028, +0.193] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0573 → 0.0830. Desired-direction raw change: +0.0257. CE +0.02; genuineness -0.06; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.033 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.054 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.018 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.285 z; 100% paired wins; Δ WER -1.39 pp
  • ES: -0.027 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE -0.00
Genuine +0.01
WER +0.00 pp
5/5 language means improve
very relaxed and unstrained
Voice dimension · vn_TENS_reg_low
Raw target direction: ↓		+0.053 z
Mean up; uncertain
Target 0.457 → 0.478
Δ score +0.021
Uncertainty and paired results

Stricter resampling range: [-0.015, +0.179] z. 2/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 1.7541 → 1.7102. Desired-direction raw change: +0.0439. CE +0.00; genuineness -0.06; WER +0.25 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.027 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.014 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.012 z; 50% paired wins; Δ WER +1.24 pp
  • NL: -0.017 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.279 z; 100% paired wins; Δ WER +0.00 pp
+0.052 z
Clearer gain
Target 0.457 → 0.477
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [+0.003, +0.137] z. 4/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 1.7541 → 1.7109. Desired-direction raw change: +0.0432. CE +0.05; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.000 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.193 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.005 z; 50% paired wins; Δ WER -1.39 pp
  • NL: +0.010 z; 75% paired wins; Δ WER +1.39 pp
  • ES: +0.052 z; 75% paired wins; Δ WER +0.00 pp
-0.019 z
Mean down / no gain
Target 0.457 → 0.450
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.098, +0.065] z. 2/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 1.7541 → 1.7699. Desired-direction raw change: -0.0158. CE -0.07; genuineness -0.18; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.123 z; 0% paired wins; Δ WER +0.00 pp
  • FR: +0.005 z; 25% paired wins; Δ WER -1.39 pp
  • NL: -0.081 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.108 z; 75% paired wins; Δ WER +0.00 pp
Try 0.75
Clearer raw-target gain on this grid		CE +0.05
Genuine +0.02
WER +0.00 pp
4/5 language means improve
very low arousal, calm or sleepy
Voice dimension · vn_AROU_reg_low
Raw target direction: ↓		-0.121 z
Mean down / no gain
Target 0.774 → 0.735
Δ score -0.039
Uncertainty and paired results

Stricter resampling range: [-0.271, +0.039] z. 1/5 language means improve. 25% of 20 paired recordings improve.

Raw head mean: 1.5699 → 1.7065. Desired-direction raw change: -0.1366. CE -0.02; genuineness -0.06; WER +0.61 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.135 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.348 z; 0% paired wins; Δ WER +0.00 pp
  • FR: -0.083 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.183 z; 0% paired wins; Δ WER +1.39 pp
  • ES: -0.127 z; 25% paired wins; Δ WER +1.67 pp
-0.025 z
Mean down / no gain
Target 0.774 → 0.766
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.108, +0.072] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.5699 → 1.5975. Desired-direction raw change: -0.0276. CE -0.01; genuineness +0.01; WER +2.05 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.083 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.143 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.004 z; 50% paired wins; Δ WER +10.60 pp
  • NL: -0.027 z; 25% paired wins; Δ WER +1.32 pp
  • ES: +0.126 z; 75% paired wins; Δ WER -1.67 pp
+0.039 z
Clearer gain
Target 0.774 → 0.785
Δ score +0.011
Uncertainty and paired results

Stricter resampling range: [+0.005, +0.081] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 1.5699 → 1.5263. Desired-direction raw change: +0.0436. CE -0.02; genuineness -0.06; WER -0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.001 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.012 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.087 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.080 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.017 z; 50% paired wins; Δ WER -1.67 pp
Try 1.0
Clearer raw-target gain on this grid
Very small scorer change; audible benefit unproven.		CE -0.02
Genuine -0.06
WER -0.33 pp
4/5 language means improve
Thankfulness Gratitude
Emotion · emo_Thankfulness_Gratitude
Raw target direction: ↑		+0.008 z
Mean up; uncertain
Target 0.415 → 0.418
Δ score +0.003
Uncertainty and paired results

Stricter resampling range: [-0.014, +0.030] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1688 → 0.1743. Desired-direction raw change: +0.0055. CE +0.02; genuineness +0.03; WER +0.59 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.020 z; 25% paired wins; Δ WER +1.56 pp
  • DE: +0.029 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.010 z; 50% paired wins; Δ WER +1.39 pp
  • NL: +0.032 z; 100% paired wins; Δ WER +0.00 pp
  • ES: -0.012 z; 25% paired wins; Δ WER +0.00 pp
-0.002 z
Mean down / no gain
Target 0.415 → 0.414
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.030, +0.018] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.1688 → 0.1676. Desired-direction raw change: -0.0012. CE -0.01; genuineness +0.10; WER +0.79 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.020 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.003 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.049 z; 25% paired wins; Δ WER +3.95 pp
  • NL: +0.018 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.000 z; 25% paired wins; Δ WER +0.00 pp
+0.024 z
Clearer gain
Target 0.415 → 0.424
Δ score +0.009
Uncertainty and paired results

Stricter resampling range: [+0.006, +0.036] z. 4/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 0.1688 → 0.1853. Desired-direction raw change: +0.0165. CE -0.03; genuineness -0.00; WER +0.79 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.039 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.022 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.030 z; 50% paired wins; Δ WER +1.32 pp
  • NL: +0.034 z; 100% paired wins; Δ WER +2.63 pp
  • ES: -0.006 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Clearer raw-target gain on this grid
Very small scorer change; audible benefit unproven.		CE -0.03
Genuine -0.00
WER +0.79 pp
4/5 language means improve
Disgust
Emotion · emo_Disgust
Raw target direction: ↑		-0.003 z
Mean down / no gain
Target 0.337 → 0.336
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.027, +0.021] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0154 → 0.0139. Desired-direction raw change: -0.0015. CE -0.01; genuineness -0.04; WER -0.02 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.001 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.031 z; 75% paired wins; Δ WER +1.56 pp
  • FR: -0.021 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.010 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.037 z; 50% paired wins; Δ WER -1.67 pp
+0.016 z
Clearer gain
Target 0.337 → 0.343
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [+0.005, +0.026] z. 5/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.0154 → 0.0231. Desired-direction raw change: +0.0076. CE +0.01; genuineness -0.03; WER -0.06 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.025 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.012 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.030 z; 75% paired wins; Δ WER +1.39 pp
  • NL: +0.010 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.001 z; 50% paired wins; Δ WER -1.67 pp
+0.006 z
Mean up; uncertain
Target 0.337 → 0.339
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.020, +0.025] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.0154 → 0.0186. Desired-direction raw change: +0.0031. CE +0.02; genuineness -0.06; WER -0.02 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.015 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.007 z; 50% paired wins; Δ WER +1.56 pp
  • FR: +0.031 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.017 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.038 z; 50% paired wins; Δ WER -1.67 pp
Try 0.75
Clearer raw-target gain on this grid
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine -0.03
WER -0.06 pp
5/5 language means improve
Longing
Emotion · emo_Longing
Raw target direction: ↑		-0.029 z
Mean down / no gain
Target 0.468 → 0.455
Δ score -0.013
Uncertainty and paired results

Stricter resampling range: [-0.125, +0.068] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.3981 → 0.3794. Desired-direction raw change: -0.0187. CE +0.00; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.022 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.123 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.114 z; 75% paired wins; Δ WER -1.32 pp
  • NL: -0.026 z; 25% paired wins; Δ WER +1.32 pp
  • ES: -0.134 z; 50% paired wins; Δ WER +0.00 pp
+0.166 z
Mean up; uncertain
Target 0.468 → 0.506
Δ score +0.038
Uncertainty and paired results

Stricter resampling range: [-0.142, +0.562] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.3981 → 0.5038. Desired-direction raw change: +0.1057. CE +0.02; genuineness -0.00; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.556 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.131 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.588 z; 100% paired wins; Δ WER -1.32 pp
  • NL: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.177 z; 25% paired wins; Δ WER +0.00 pp
+0.009 z
Mean up; uncertain
Target 0.468 → 0.468
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.142, +0.149] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.3981 → 0.4041. Desired-direction raw change: +0.0060. CE +0.00; genuineness -0.09; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.081 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.013 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.217 z; 75% paired wins; Δ WER -1.32 pp
  • NL: -0.033 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.230 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain		CE +0.02
Genuine -0.00
WER -0.26 pp
2/5 language means improve
strongly negative vocal affect
Voice dimension · vn_VALN_reg_low
Raw target direction: ↓		+0.127 z
Mean up; uncertain
Target 0.229 → 0.254
Δ score +0.025
Uncertainty and paired results

Stricter resampling range: [-0.100, +0.370] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 3.0900 → 2.9471. Desired-direction raw change: +0.1429. CE +0.00; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.074 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.109 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.318 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.098 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.447 z; 100% paired wins; Δ WER +0.00 pp
+0.153 z
Mean up; uncertain
Target 0.229 → 0.262
Δ score +0.033
Uncertainty and paired results

Stricter resampling range: [-0.056, +0.408] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.0900 → 2.9171. Desired-direction raw change: +0.1729. CE -0.05; genuineness -0.09; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.106 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.043 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.296 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.004 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.536 z; 100% paired wins; Δ WER +0.00 pp
+0.105 z
Mean up; uncertain
Target 0.229 → 0.258
Δ score +0.028
Uncertainty and paired results

Stricter resampling range: [-0.002, +0.238] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 3.0900 → 2.9718. Desired-direction raw change: +0.1182. CE +0.02; genuineness -0.12; WER +2.70 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.016 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.027 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.166 z; 75% paired wins; Δ WER +6.58 pp
  • NL: +0.065 z; 75% paired wins; Δ WER +5.26 pp
  • ES: +0.304 z; 50% paired wins; Δ WER +1.67 pp
Try 0.75
Positive raw mean; uncertain		CE -0.05
Genuine -0.09
WER +0.00 pp
3/5 language means improve
Doubt
Emotion · emo_Doubt
Raw target direction: ↑		+0.139 z
Mean up; uncertain
Target 0.440 → 0.472
Δ score +0.033
Uncertainty and paired results

Stricter resampling range: [-0.018, +0.343] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.2784 → 0.3454. Desired-direction raw change: +0.0671. CE -0.02; genuineness -0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.000 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.257 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.044 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.439 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.042 z; 50% paired wins; Δ WER +0.00 pp
+0.107 z
Mean up; uncertain
Target 0.440 → 0.467
Δ score +0.027
Uncertainty and paired results

Stricter resampling range: [-0.037, +0.320] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.2784 → 0.3303. Desired-direction raw change: +0.0519. CE -0.04; genuineness -0.05; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.134 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.052 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.033 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.445 z; 75% paired wins; Δ WER +1.39 pp
  • ES: +0.043 z; 50% paired wins; Δ WER +0.00 pp
+0.073 z
Mean up; uncertain
Target 0.440 → 0.434
Δ score -0.006

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.166, +0.292] z. 4/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.2784 → 0.3137. Desired-direction raw change: +0.0353. CE -0.00; genuineness -0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.041 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.235 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.290 z; 0% paired wins; Δ WER +0.00 pp
  • NL: +0.378 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.001 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain		CE -0.02
Genuine -0.00
WER +0.00 pp
4/5 language means improve
Strong Formal Style
Talking style · vn_S_FORM_reg
Raw target direction: ↑		+0.023 z
Mean up; uncertain
Target 0.869 → 0.877
Δ score +0.008
Uncertainty and paired results

Stricter resampling range: [-0.244, +0.357] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.5639 → 2.5872. Desired-direction raw change: +0.0232. CE +0.01; genuineness -0.02; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.181 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.519 z; 100% paired wins; Δ WER +1.32 pp
  • FR: -0.291 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.081 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.213 z; 25% paired wins; Δ WER +0.00 pp
-0.058 z
Mean down / no gain
Target 0.869 → 0.873
Δ score +0.004

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.326, +0.237] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.5639 → 2.5055. Desired-direction raw change: -0.0584. CE -0.02; genuineness -0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.145 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.401 z; 100% paired wins; Δ WER +0.00 pp
  • FR: -0.231 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.104 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.418 z; 0% paired wins; Δ WER +0.00 pp
+0.128 z
Mean up; uncertain
Target 0.869 → 0.905
Δ score +0.036
Uncertainty and paired results

Stricter resampling range: [-0.145, +0.447] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.5639 → 2.6932. Desired-direction raw change: +0.1293. CE +0.02; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.033 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.495 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.040 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.436 z; 100% paired wins; Δ WER +0.00 pp
  • ES: -0.217 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE +0.02
Genuine +0.01
WER +0.00 pp
2/5 language means improve
Very formal, not casual
Style contrast · formal_not_casual
Raw target direction: ↑		+0.117 z
Mean up; uncertain
Target 0.897 → 0.911
Δ score +0.014
Uncertainty and paired results

Stricter resampling range: [-0.051, +0.292] z. 4/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 2.6377 → 2.7555. Desired-direction raw change: +0.1178. CE +0.04; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.050 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.129 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.159 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.145 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.391 z; 100% paired wins; Δ WER +0.00 pp

Opposite head vn_S_CASU_reg: desired-direction change +0.049 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

+0.076 z
Mean up; uncertain
Target 0.897 → 0.907
Δ score +0.010
Uncertainty and paired results

Stricter resampling range: [+0.012, +0.125] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.6377 → 2.7144. Desired-direction raw change: +0.0767. CE +0.02; genuineness -0.08; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.114 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.055 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.093 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.020 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.139 z; 50% paired wins; Δ WER +0.00 pp

Opposite head vn_S_CASU_reg: desired-direction change +0.066 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

+0.108 z
Mean up; uncertain
Target 0.897 → 0.910
Δ score +0.013
Uncertainty and paired results

Stricter resampling range: [-0.040, +0.300] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.6377 → 2.7464. Desired-direction raw change: +0.1087. CE +0.03; genuineness -0.05; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.130 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.074 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.099 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.029 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.414 z; 100% paired wins; Δ WER +0.00 pp

Opposite head vn_S_CASU_reg: desired-direction change +0.132 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

Try 0.5
Positive raw mean; uncertain
Primary and opposite heads reported in cell details.		CE +0.04
Genuine -0.03
WER +0.00 pp
4/5 language means improve
Relief
Emotion · emo_Relief
Raw target direction: ↑		+0.115 z
Mean up; uncertain
Target 0.527 → 0.567
Δ score +0.040
Uncertainty and paired results

Stricter resampling range: [-0.011, +0.246] z. 4/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 0.4253 → 0.5054. Desired-direction raw change: +0.0802. CE -0.01; genuineness -0.07; WER -0.54 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.298 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.097 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.212 z; 75% paired wins; Δ WER -1.39 pp
  • NL: -0.044 z; 25% paired wins; Δ WER -1.32 pp
  • ES: +0.012 z; 50% paired wins; Δ WER +0.00 pp
+0.106 z
Mean up; uncertain
Target 0.527 → 0.558
Δ score +0.031
Uncertainty and paired results

Stricter resampling range: [-0.062, +0.294] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.4253 → 0.4990. Desired-direction raw change: +0.0738. CE +0.01; genuineness -0.08; WER +1.83 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.252 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.019 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.363 z; 100% paired wins; Δ WER -0.07 pp
  • NL: +0.007 z; 50% paired wins; Δ WER +9.21 pp
  • ES: -0.112 z; 25% paired wins; Δ WER +0.00 pp
-0.029 z
Mean down / no gain
Target 0.527 → 0.518
Δ score -0.009
Uncertainty and paired results

Stricter resampling range: [-0.153, +0.074] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.4253 → 0.4053. Desired-direction raw change: -0.0199. CE -0.06; genuineness -0.12; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.121 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.031 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.025 z; 25% paired wins; Δ WER -1.39 pp
  • NL: -0.048 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.223 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain		CE -0.01
Genuine -0.07
WER -0.54 pp
4/5 language means improve
Concentration
Emotion · emo_Concentration
Raw target direction: ↑		-0.030 z
Mean down / no gain
Target 0.941 → 0.932
Δ score -0.008
Uncertainty and paired results

Stricter resampling range: [-0.144, +0.057] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.8646 → 1.8425. Desired-direction raw change: -0.0221. CE +0.02; genuineness -0.03; WER -0.87 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.232 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.001 z; 50% paired wins; Δ WER -1.32 pp
  • FR: +0.096 z; 75% paired wins; Δ WER -1.39 pp
  • NL: -0.002 z; 50% paired wins; Δ WER +1.32 pp
  • ES: -0.012 z; 25% paired wins; Δ WER -2.94 pp
+0.008 z
Mean up; uncertain
Target 0.941 → 0.931
Δ score -0.009

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.162, +0.207] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 1.8646 → 1.8708. Desired-direction raw change: +0.0061. CE -0.01; genuineness -0.00; WER -1.13 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.253 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.000 z; 75% paired wins; Δ WER -1.32 pp
  • FR: +0.004 z; 50% paired wins; Δ WER -1.39 pp
  • NL: +0.342 z; 100% paired wins; Δ WER +0.00 pp
  • ES: -0.051 z; 50% paired wins; Δ WER -2.94 pp
+0.100 z
Mean up; uncertain
Target 0.941 → 0.943
Δ score +0.003
Uncertainty and paired results

Stricter resampling range: [-0.030, +0.234] z. 3/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 1.8646 → 1.9377. Desired-direction raw change: +0.0731. CE -0.01; genuineness -0.09; WER -1.13 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.062 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.102 z; 75% paired wins; Δ WER -1.32 pp
  • FR: +0.230 z; 100% paired wins; Δ WER -1.39 pp
  • NL: +0.248 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.018 z; 50% paired wins; Δ WER -2.94 pp
Try 1.0
Positive raw mean; uncertain		CE -0.01
Genuine -0.09
WER -1.13 pp
3/5 language means improve
emotionally vulnerable and exposed
Voice dimension · vn_VULN_reg_high
Raw target direction: ↑		-0.045 z
Mean down / no gain
Target 0.413 → 0.404
Δ score -0.009
Uncertainty and paired results

Stricter resampling range: [-0.119, +0.051] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.4934 → 2.4488. Desired-direction raw change: -0.0446. CE +0.01; genuineness -0.03; WER +0.54 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.142 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.011 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.090 z; 25% paired wins; Δ WER +2.63 pp
  • NL: -0.083 z; 50% paired wins; Δ WER -1.39 pp
  • ES: +0.100 z; 50% paired wins; Δ WER +1.47 pp
+0.008 z
Mean up; uncertain
Target 0.413 → 0.423
Δ score +0.010
Uncertainty and paired results

Stricter resampling range: [-0.203, +0.222] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.4934 → 2.5015. Desired-direction raw change: +0.0081. CE +0.03; genuineness -0.04; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.027 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.111 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.165 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.292 z; 50% paired wins; Δ WER -1.39 pp
  • ES: +0.307 z; 100% paired wins; Δ WER +0.00 pp
+0.089 z
Mean up; uncertain
Target 0.413 → 0.453
Δ score +0.040
Uncertainty and paired results

Stricter resampling range: [-0.091, +0.251] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.4934 → 2.5817. Desired-direction raw change: +0.0883. CE +0.01; genuineness -0.03; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.072 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.110 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.200 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.125 z; 25% paired wins; Δ WER -1.39 pp
  • ES: +0.339 z; 75% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE +0.01
Genuine -0.03
WER -0.28 pp
4/5 language means improve
Awe
Emotion · emo_Awe
Raw target direction: ↑		+0.087 z
Mean up; uncertain
Target 0.623 → 0.657
Δ score +0.034
Uncertainty and paired results

Stricter resampling range: [-0.106, +0.249] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.3321 → 0.3684. Desired-direction raw change: +0.0363. CE -0.01; genuineness -0.08; WER -0.62 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.212 z; 50% paired wins; Δ WER -1.56 pp
  • DE: +0.319 z; 50% paired wins; Δ WER -1.56 pp
  • FR: +0.202 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.082 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.043 z; 50% paired wins; Δ WER +0.00 pp
+0.055 z
Mean up; uncertain
Target 0.623 → 0.654
Δ score +0.030
Uncertainty and paired results

Stricter resampling range: [-0.146, +0.364] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.3321 → 0.3550. Desired-direction raw change: +0.0230. CE -0.01; genuineness -0.09; WER -0.62 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.113 z; 50% paired wins; Δ WER -1.56 pp
  • DE: -0.182 z; 50% paired wins; Δ WER -1.56 pp
  • FR: +0.096 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.069 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.542 z; 100% paired wins; Δ WER +0.00 pp
+0.072 z
Mean up; uncertain
Target 0.623 → 0.658
Δ score +0.034
Uncertainty and paired results

Stricter resampling range: [-0.175, +0.247] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.3321 → 0.3624. Desired-direction raw change: +0.0304. CE +0.01; genuineness -0.08; WER -0.62 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.289 z; 25% paired wins; Δ WER -1.56 pp
  • DE: +0.218 z; 75% paired wins; Δ WER -1.56 pp
  • FR: +0.183 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.028 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.278 z; 50% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain		CE -0.01
Genuine -0.08
WER -0.62 pp
4/5 language means improve
Strong introspective monologue delivery, not monotone speech
Talking style · vn_S_MONO_reg
Raw target direction: ↑		+0.081 z
Mean up; uncertain
Target 0.713 → 0.741
Δ score +0.028
Uncertainty and paired results

Stricter resampling range: [-0.103, +0.249] z. 4/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 3.4079 → 3.5115. Desired-direction raw change: +0.1036. CE -0.03; genuineness -0.03; WER -1.05 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.323 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.057 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.200 z; 100% paired wins; Δ WER -5.26 pp
  • NL: +0.028 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.200 z; 0% paired wins; Δ WER +0.00 pp
+0.018 z
Mean up; uncertain
Target 0.713 → 0.721
Δ score +0.008
Uncertainty and paired results

Stricter resampling range: [-0.065, +0.167] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 3.4079 → 3.4307. Desired-direction raw change: +0.0228. CE +0.03; genuineness -0.01; WER -0.77 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.304 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.073 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.060 z; 50% paired wins; Δ WER -5.26 pp
  • NL: -0.042 z; 25% paired wins; Δ WER +1.39 pp
  • ES: -0.039 z; 50% paired wins; Δ WER +0.00 pp
+0.069 z
Mean up; uncertain
Target 0.713 → 0.734
Δ score +0.021
Uncertainty and paired results

Stricter resampling range: [-0.062, +0.218] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 3.4079 → 3.4957. Desired-direction raw change: +0.0878. CE +0.01; genuineness +0.02; WER -0.72 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.220 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.011 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.217 z; 75% paired wins; Δ WER -5.26 pp
  • NL: -0.110 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.028 z; 50% paired wins; Δ WER +1.67 pp
Try 0.5
Positive raw mean; uncertain		CE -0.03
Genuine -0.03
WER -1.05 pp
4/5 language means improve
very tense and strained
Voice dimension · vn_TENS_reg_high
Raw target direction: ↑		-0.006 z
Mean down / no gain
Target 0.794 → 0.798
Δ score +0.004

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.216, +0.202] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.4711 → 2.4658. Desired-direction raw change: -0.0053. CE +0.03; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.252 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.252 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.175 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.000 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.208 z; 25% paired wins; Δ WER +0.00 pp
+0.070 z
Mean up; uncertain
Target 0.794 → 0.811
Δ score +0.017
Uncertainty and paired results

Stricter resampling range: [-0.280, +0.377] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.4711 → 2.5294. Desired-direction raw change: +0.0582. CE -0.00; genuineness -0.09; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.211 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.459 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.301 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.457 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.161 z; 25% paired wins; Δ WER +0.00 pp
+0.078 z
Mean up; uncertain
Target 0.794 → 0.791
Δ score -0.002

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.226, +0.288] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.4711 → 2.5362. Desired-direction raw change: +0.0651. CE -0.02; genuineness -0.08; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.069 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.225 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.348 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.172 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.423 z; 25% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE -0.02
Genuine -0.08
WER +0.00 pp
4/5 language means improve
Fatigue Exhaustion
Emotion · emo_Fatigue_Exhaustion
Raw target direction: ↑		+0.045 z
Mean up; uncertain
Target 0.687 → 0.702
Δ score +0.015
Uncertainty and paired results

Stricter resampling range: [-0.401, +0.610] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.6722 → 0.6967. Desired-direction raw change: +0.0245. CE -0.01; genuineness +0.02; WER -0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.093 z; 25% paired wins; Δ WER -1.56 pp
  • DE: +0.276 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.872 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.436 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.392 z; 25% paired wins; Δ WER +0.00 pp
-0.053 z
Mean down / no gain
Target 0.687 → 0.686
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.324, +0.274] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.6722 → 0.6437. Desired-direction raw change: -0.0285. CE +0.03; genuineness +0.02; WER -0.03 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.214 z; 0% paired wins; Δ WER -1.56 pp
  • DE: +0.271 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.287 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.207 z; 50% paired wins; Δ WER +1.39 pp
  • ES: -0.400 z; 25% paired wins; Δ WER +0.00 pp
+0.076 z
Mean up; uncertain
Target 0.687 → 0.710
Δ score +0.023
Uncertainty and paired results

Stricter resampling range: [-0.134, +0.338] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.6722 → 0.7133. Desired-direction raw change: +0.0411. CE -0.02; genuineness -0.04; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.179 z; 25% paired wins; Δ WER -1.56 pp
  • DE: +0.226 z; 75% paired wins; Δ WER +1.56 pp
  • FR: +0.468 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.041 z; 50% paired wins; Δ WER +1.39 pp
  • ES: -0.093 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE -0.02
Genuine -0.04
WER +0.28 pp
2/5 language means improve
Strong Cartoonish Style
Talking style · vn_S_CART_reg
Raw target direction: ↑		+0.004 z
Mean up; uncertain
Target 0.262 → 0.264
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.093, +0.108] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.7940 → 0.7994. Desired-direction raw change: +0.0053. CE -0.03; genuineness -0.11; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.079 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.092 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.019 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.151 z; 75% paired wins; Δ WER -1.32 pp
  • ES: -0.100 z; 0% paired wins; Δ WER +0.00 pp
+0.060 z
Mean up; uncertain
Target 0.262 → 0.281
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [-0.105, +0.277] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.7940 → 0.8763. Desired-direction raw change: +0.0822. CE +0.02; genuineness -0.08; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.084 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.009 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.136 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.380 z; 100% paired wins; Δ WER -1.32 pp
  • ES: +0.149 z; 75% paired wins; Δ WER +0.00 pp
+0.007 z
Mean up; uncertain
Target 0.262 → 0.260
Δ score -0.001

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.063, +0.075] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.7940 → 0.8034. Desired-direction raw change: +0.0094. CE -0.03; genuineness -0.02; WER +0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.100 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.106 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.042 z; 50% paired wins; Δ WER +1.32 pp
  • NL: +0.011 z; 50% paired wins; Δ WER -1.32 pp
  • ES: -0.025 z; 50% paired wins; Δ WER +1.67 pp
Try 0.75
Positive raw mean; uncertain		CE +0.02
Genuine -0.08
WER -0.26 pp
2/5 language means improve
strongly positive vocal affect
Voice dimension · vn_VALN_reg_high
Raw target direction: ↑		+0.058 z
Mean up; uncertain
Target 0.865 → 0.876
Δ score +0.012
Uncertainty and paired results

Stricter resampling range: [-0.052, +0.140] z. 4/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 3.5437 → 3.6098. Desired-direction raw change: +0.0660. CE -0.06; genuineness +0.00; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.081 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.145 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.117 z; 50% paired wins; Δ WER +1.32 pp
  • NL: +0.139 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.045 z; 50% paired wins; Δ WER +0.00 pp
-0.005 z
Mean down / no gain
Target 0.865 → 0.858
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.098, +0.088] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.5437 → 3.5380. Desired-direction raw change: -0.0057. CE -0.06; genuineness -0.03; WER +0.59 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.109 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.059 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.136 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.022 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.082 z; 75% paired wins; Δ WER +2.94 pp
-0.044 z
Mean down / no gain
Target 0.865 → 0.841
Δ score -0.024
Uncertainty and paired results

Stricter resampling range: [-0.094, +0.019] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.5437 → 3.4940. Desired-direction raw change: -0.0498. CE -0.03; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.051 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.072 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.014 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.078 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.107 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain		CE -0.06
Genuine +0.00
WER +0.26 pp
4/5 language means improve
highly aroused and activated
Voice dimension · vn_AROU_reg_high
Raw target direction: ↑		+0.045 z
Mean up; uncertain
Target 0.472 → 0.489
Δ score +0.017
Uncertainty and paired results

Stricter resampling range: [-0.118, +0.216] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.3577 → 2.4085. Desired-direction raw change: +0.0508. CE -0.02; genuineness -0.00; WER -0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.079 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.070 z; 50% paired wins; Δ WER -1.56 pp
  • FR: -0.176 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.299 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.110 z; 75% paired wins; Δ WER +0.00 pp
-0.043 z
Mean down / no gain
Target 0.472 → 0.454
Δ score -0.019
Uncertainty and paired results

Stricter resampling range: [-0.158, +0.111] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.3577 → 2.3098. Desired-direction raw change: -0.0479. CE -0.06; genuineness -0.08; WER +0.02 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.104 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.100 z; 25% paired wins; Δ WER -1.56 pp
  • FR: -0.001 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.187 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.195 z; 25% paired wins; Δ WER +1.67 pp
+0.057 z
Mean up; uncertain
Target 0.472 → 0.494
Δ score +0.021
Uncertainty and paired results

Stricter resampling range: [-0.141, +0.246] z. 4/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.3577 → 2.4219. Desired-direction raw change: +0.0642. CE -0.03; genuineness -0.03; WER -0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.265 z; 0% paired wins; Δ WER +0.00 pp
  • DE: +0.102 z; 50% paired wins; Δ WER -1.56 pp
  • FR: +0.029 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.354 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.066 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE -0.03
Genuine -0.03
WER -0.31 pp
4/5 language means improve
Interest
Emotion · emo_Interest
Raw target direction: ↑		-0.036 z
Mean down / no gain
Target 0.764 → 0.759
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.129, +0.046] z. 3/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 2.1408 → 2.1083. Desired-direction raw change: -0.0325. CE +0.00; genuineness -0.04; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.122 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.049 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.045 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.007 z; 25% paired wins; Δ WER +1.32 pp
  • ES: -0.157 z; 0% paired wins; Δ WER +0.00 pp
+0.057 z
Mean up; uncertain
Target 0.764 → 0.784
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [-0.030, +0.145] z. 4/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 2.1408 → 2.1926. Desired-direction raw change: +0.0518. CE -0.01; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.004 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.178 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.124 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.043 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.065 z; 25% paired wins; Δ WER +0.00 pp
-0.025 z
Mean down / no gain
Target 0.764 → 0.758
Δ score -0.005
Uncertainty and paired results

Stricter resampling range: [-0.112, +0.079] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.1408 → 2.1183. Desired-direction raw change: -0.0225. CE +0.03; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.076 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.136 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.038 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.004 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.149 z; 25% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain		CE -0.01
Genuine -0.03
WER +0.00 pp
4/5 language means improve
Malevolence Malice
Emotion · emo_Malevolence_Malice
Raw target direction: ↑		+0.003 z
Mean up; uncertain
Target 0.312 → 0.314
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.051, +0.069] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.0219 → 0.0233. Desired-direction raw change: +0.0014. CE -0.01; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.001 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.113 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.019 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.005 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.078 z; 25% paired wins; Δ WER +0.00 pp
-0.030 z
Mean down / no gain
Target 0.312 → 0.301
Δ score -0.011
Uncertainty and paired results

Stricter resampling range: [-0.094, +0.027] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.0219 → 0.0062. Desired-direction raw change: -0.0157. CE -0.02; genuineness -0.03; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.015 z; 25% paired wins; Δ WER +1.39 pp
  • DE: -0.129 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.007 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.047 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.060 z; 25% paired wins; Δ WER +0.00 pp
+0.056 z
Mean up; uncertain
Target 0.312 → 0.331
Δ score +0.019
Uncertainty and paired results

Stricter resampling range: [-0.013, +0.159] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0219 → 0.0511. Desired-direction raw change: +0.0293. CE +0.00; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.030 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.243 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.007 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.034 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.035 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain		CE +0.00
Genuine -0.07
WER +0.00 pp
4/5 language means improve
very warm vocal delivery
Voice dimension · vn_WARM_reg_high
Raw target direction: ↑		-0.042 z
Mean down / no gain
Target 0.874 → 0.869
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.155, +0.050] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.6782 → 3.6404. Desired-direction raw change: -0.0378. CE +0.01; genuineness +0.03; WER -0.61 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.010 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.045 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.229 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.018 z; 50% paired wins; Δ WER -1.39 pp
  • ES: +0.092 z; 75% paired wins; Δ WER -1.67 pp
+0.055 z
Mean up; uncertain
Target 0.874 → 0.884
Δ score +0.010
Uncertainty and paired results

Stricter resampling range: [-0.039, +0.171] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 3.6782 → 3.7282. Desired-direction raw change: +0.0500. CE -0.02; genuineness +0.04; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.072 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.023 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.082 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.026 z; 75% paired wins; Δ WER -1.39 pp
  • ES: +0.236 z; 75% paired wins; Δ WER +0.00 pp
+0.017 z
Mean up; uncertain
Target 0.874 → 0.879
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.110, +0.123] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 3.6782 → 3.6938. Desired-direction raw change: +0.0156. CE -0.01; genuineness +0.00; WER -0.61 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.081 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.050 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.176 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.078 z; 75% paired wins; Δ WER -1.39 pp
  • ES: +0.152 z; 100% paired wins; Δ WER -1.67 pp
Try 0.75
Positive raw mean; uncertain		CE -0.02
Genuine +0.04
WER -0.28 pp
4/5 language means improve
Elation
Emotion · emo_Elation
Raw target direction: ↑		-0.048 z
Mean down / no gain
Target 0.366 → 0.349
Δ score -0.018
Uncertainty and paired results

Stricter resampling range: [-0.107, +0.020] z. 1/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.2010 → 0.1668. Desired-direction raw change: -0.0342. CE -0.01; genuineness +0.01; WER -0.83 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.141 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.049 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.065 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.060 z; 25% paired wins; Δ WER -4.17 pp
  • ES: -0.053 z; 25% paired wins; Δ WER +0.00 pp
+0.051 z
Mean up; uncertain
Target 0.366 → 0.386
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [-0.026, +0.155] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.2010 → 0.2376. Desired-direction raw change: +0.0366. CE -0.03; genuineness -0.01; WER -0.83 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.038 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.024 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.057 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.039 z; 50% paired wins; Δ WER -4.17 pp
  • ES: +0.220 z; 100% paired wins; Δ WER +0.00 pp
-0.049 z
Mean down / no gain
Target 0.366 → 0.348
Δ score -0.019
Uncertainty and paired results

Stricter resampling range: [-0.131, +0.026] z. 2/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.2010 → 0.1657. Desired-direction raw change: -0.0353. CE -0.02; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.166 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.108 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.000 z; 25% paired wins; Δ WER +4.17 pp
  • NL: +0.051 z; 50% paired wins; Δ WER -4.17 pp
  • ES: -0.024 z; 25% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain		CE -0.03
Genuine -0.01
WER -0.83 pp
3/5 language means improve
Helplessness
Emotion · emo_Helplessness
Raw target direction: ↑		+0.051 z
Mean up; uncertain
Target 0.359 → 0.378
Δ score +0.020
Uncertainty and paired results

Stricter resampling range: [-0.011, +0.133] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.1723 → 0.2083. Desired-direction raw change: +0.0360. CE +0.02; genuineness +0.04; WER -2.11 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.013 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.031 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.176 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.070 z; 75% paired wins; Δ WER -10.53 pp
  • ES: +0.026 z; 75% paired wins; Δ WER +0.00 pp
-0.032 z
Mean down / no gain
Target 0.359 → 0.347
Δ score -0.012
Uncertainty and paired results

Stricter resampling range: [-0.081, +0.021] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.1723 → 0.1495. Desired-direction raw change: -0.0227. CE +0.03; genuineness +0.02; WER -2.11 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.050 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.058 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.096 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.058 z; 50% paired wins; Δ WER -10.53 pp
  • ES: +0.001 z; 50% paired wins; Δ WER +0.00 pp
-0.022 z
Mean down / no gain
Target 0.359 → 0.351
Δ score -0.008
Uncertainty and paired results

Stricter resampling range: [-0.060, +0.021] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.1723 → 0.1565. Desired-direction raw change: -0.0158. CE -0.01; genuineness -0.08; WER -2.11 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.002 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.060 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.060 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.041 z; 75% paired wins; Δ WER -10.53 pp
  • ES: -0.030 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain		CE +0.02
Genuine +0.04
WER -2.11 pp
4/5 language means improve
Strong Whisper-Talk Style
Talking style · vn_S_WHIS_reg
Raw target direction: ↑		+0.047 z
Mean up; uncertain
Target 0.294 → 0.310
Δ score +0.016
Uncertainty and paired results

Stricter resampling range: [-0.000, +0.103] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.8636 → 0.9389. Desired-direction raw change: +0.0752. CE -0.04; genuineness +0.02; WER +1.14 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.135 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.057 z; 50% paired wins; Δ WER +2.63 pp
  • FR: +0.048 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.019 z; 50% paired wins; Δ WER +1.39 pp
  • ES: -0.022 z; 25% paired wins; Δ WER +1.67 pp
-0.035 z
Mean down / no gain
Target 0.294 → 0.286
Δ score -0.008
Uncertainty and paired results

Stricter resampling range: [-0.122, +0.107] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.8636 → 0.8080. Desired-direction raw change: -0.0556. CE -0.03; genuineness +0.01; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.037 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.124 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.203 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.122 z; 25% paired wins; Δ WER +1.39 pp
  • ES: -0.093 z; 25% paired wins; Δ WER +0.00 pp
-0.097 z
Mean down / no gain
Target 0.294 → 0.262
Δ score -0.032
Uncertainty and paired results

Stricter resampling range: [-0.123, -0.061] z. 0/5 language means improve. 25% of 20 paired recordings improve.

Raw head mean: 0.8636 → 0.7078. Desired-direction raw change: -0.1558. CE -0.02; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.125 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.123 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.108 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.092 z; 0% paired wins; Δ WER +0.00 pp
  • ES: -0.040 z; 25% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.04
Genuine +0.02
WER +1.14 pp
4/5 language means improve
Distress
Emotion · emo_Distress
Raw target direction: ↑		-0.009 z
Mean down / no gain
Target 0.362 → 0.358
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.038, +0.013] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1684 → 0.1613. Desired-direction raw change: -0.0071. CE -0.03; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.010 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.014 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.013 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.037 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.043 z; 25% paired wins; Δ WER +0.00 pp
+0.016 z
Mean up; uncertain
Target 0.362 → 0.368
Δ score +0.006
Uncertainty and paired results

Stricter resampling range: [-0.050, +0.085] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.1684 → 0.1815. Desired-direction raw change: +0.0131. CE +0.02; genuineness -0.02; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.050 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.019 z; 50% paired wins; Δ WER +1.32 pp
  • FR: +0.044 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.120 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.052 z; 25% paired wins; Δ WER +0.00 pp
+0.033 z
Mean up; uncertain
Target 0.362 → 0.375
Δ score +0.013
Uncertainty and paired results

Stricter resampling range: [-0.020, +0.088] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1684 → 0.1958. Desired-direction raw change: +0.0274. CE -0.02; genuineness -0.04; WER +4.64 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.017 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.050 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.030 z; 75% paired wins; Δ WER +2.63 pp
  • NL: +0.119 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.049 z; 25% paired wins; Δ WER +20.59 pp
Try 1.0
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.02
Genuine -0.04
WER +4.64 pp
4/5 language means improve
Hope Enthusiasm Optimism
Emotion · emo_Hope_Enthusiasm_Optimism
Raw target direction: ↑		-0.019 z
Mean down / no gain
Target 0.804 → 0.805
Δ score +0.001

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.165, +0.116] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 1.3521 → 1.3378. Desired-direction raw change: -0.0143. CE -0.02; genuineness +0.01; WER -0.56 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.248 z; 0% paired wins; Δ WER +0.00 pp
  • DE: +0.167 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.070 z; 75% paired wins; Δ WER +1.39 pp
  • NL: -0.092 z; 50% paired wins; Δ WER -4.17 pp
  • ES: +0.009 z; 50% paired wins; Δ WER +0.00 pp
+0.032 z
Mean up; uncertain
Target 0.804 → 0.816
Δ score +0.012
Uncertainty and paired results

Stricter resampling range: [-0.109, +0.153] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.3521 → 1.3764. Desired-direction raw change: +0.0243. CE -0.01; genuineness -0.01; WER -0.83 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.183 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.110 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.186 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.052 z; 25% paired wins; Δ WER -4.17 pp
  • ES: +0.099 z; 50% paired wins; Δ WER +0.00 pp
-0.075 z
Mean down / no gain
Target 0.804 → 0.782
Δ score -0.022
Uncertainty and paired results

Stricter resampling range: [-0.155, -0.024] z. 0/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.3521 → 1.2954. Desired-direction raw change: -0.0567. CE -0.00; genuineness +0.02; WER -0.52 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.057 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.041 z; 75% paired wins; Δ WER +1.56 pp
  • FR: -0.045 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.011 z; 50% paired wins; Δ WER -4.17 pp
  • ES: -0.220 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.01
Genuine -0.01
WER -0.83 pp
3/5 language means improve
Teasing
Emotion · emo_Teasing
Raw target direction: ↑		+0.025 z
Mean up; uncertain
Target 0.452 → 0.451
Δ score -0.001

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.178, +0.210] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.2895 → 0.3027. Desired-direction raw change: +0.0132. CE +0.03; genuineness +0.06; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.081 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.283 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.043 z; 75% paired wins; Δ WER -1.39 pp
  • NL: +0.282 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.162 z; 50% paired wins; Δ WER +0.00 pp
-0.141 z
Mean down / no gain
Target 0.452 → 0.402
Δ score -0.049
Uncertainty and paired results

Stricter resampling range: [-0.312, +0.011] z. 1/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.2895 → 0.2134. Desired-direction raw change: -0.0761. CE -0.02; genuineness +0.02; WER +0.60 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.086 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.328 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.049 z; 50% paired wins; Δ WER -1.39 pp
  • NL: -0.308 z; 25% paired wins; Δ WER +2.70 pp
  • ES: +0.064 z; 50% paired wins; Δ WER +1.67 pp
-0.111 z
Mean down / no gain
Target 0.452 → 0.419
Δ score -0.032
Uncertainty and paired results

Stricter resampling range: [-0.223, -0.020] z. 0/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.2895 → 0.2299. Desired-direction raw change: -0.0596. CE +0.01; genuineness +0.03; WER +0.25 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.080 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.277 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.029 z; 25% paired wins; Δ WER +1.24 pp
  • NL: -0.162 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.005 z; 50% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.03
Genuine +0.06
WER -0.28 pp
3/5 language means improve
very slow speaking tempo
Voice dimension · vn_TEMP_reg_low
Raw target direction: ↓		-0.018 z
Mean down / no gain
Target 0.492 → 0.485
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.066, +0.042] z. 1/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 2.1389 → 2.1610. Desired-direction raw change: -0.0221. CE -0.01; genuineness +0.06; WER +0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.088 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.086 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.028 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.037 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.025 z; 50% paired wins; Δ WER +1.67 pp
+0.023 z
Mean up; uncertain
Target 0.492 → 0.501
Δ score +0.009
Uncertainty and paired results

Stricter resampling range: [-0.013, +0.050] z. 4/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.1389 → 2.1111. Desired-direction raw change: +0.0278. CE +0.01; genuineness -0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.049 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.020 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.055 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.026 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.034 z; 50% paired wins; Δ WER +0.00 pp
+0.005 z
Mean up; uncertain
Target 0.492 → 0.494
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.065, +0.059] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.1389 → 2.1327. Desired-direction raw change: +0.0062. CE +0.01; genuineness +0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.107 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.038 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.081 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.002 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.016 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine -0.00
WER +0.00 pp
4/5 language means improve
Strong Ranting/Angry Style
Talking style · vn_S_RANT_reg
Raw target direction: ↑		+0.023 z
Mean up; uncertain
Target 0.163 → 0.167
Δ score +0.004
Uncertainty and paired results

Stricter resampling range: [-0.050, +0.103] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0940 → 0.1236. Desired-direction raw change: +0.0295. CE +0.02; genuineness +0.04; WER -0.60 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.148 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.002 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.009 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.081 z; 25% paired wins; Δ WER -1.32 pp
  • ES: +0.059 z; 50% paired wins; Δ WER -1.67 pp
-0.007 z
Mean down / no gain
Target 0.163 → 0.155
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.158, +0.118] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0940 → 0.0845. Desired-direction raw change: -0.0095. CE +0.02; genuineness +0.01; WER +0.03 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.099 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.146 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.122 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.212 z; 25% paired wins; Δ WER -1.32 pp
  • ES: +0.053 z; 50% paired wins; Δ WER +1.47 pp
-0.060 z
Mean down / no gain
Target 0.163 → 0.143
Δ score -0.020
Uncertainty and paired results

Stricter resampling range: [-0.178, +0.017] z. 1/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0940 → 0.0170. Desired-direction raw change: -0.0771. CE +0.04; genuineness +0.08; WER +0.51 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.041 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.015 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.021 z; 75% paired wins; Δ WER +2.78 pp
  • NL: -0.276 z; 25% paired wins; Δ WER +1.46 pp
  • ES: -0.031 z; 50% paired wins; Δ WER -1.67 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.02
Genuine +0.04
WER -0.60 pp
2/5 language means improve
Emotional Numbness
Emotion · emo_Emotional_Numbness
Raw target direction: ↑		+0.006 z
Mean up; uncertain
Target 0.359 → 0.361
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.015, +0.024] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.1128 → 0.1183. Desired-direction raw change: +0.0055. CE +0.01; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.011 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.006 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.025 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.025 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.024 z; 75% paired wins; Δ WER +0.00 pp
+0.022 z
Mean up; uncertain
Target 0.359 → 0.367
Δ score +0.008
Uncertainty and paired results

Stricter resampling range: [-0.017, +0.085] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1128 → 0.1340. Desired-direction raw change: +0.0213. CE -0.02; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.021 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.010 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.121 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.010 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.031 z; 50% paired wins; Δ WER +0.00 pp
+0.005 z
Mean up; uncertain
Target 0.359 → 0.360
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.035, +0.044] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.1128 → 0.1171. Desired-direction raw change: +0.0044. CE -0.01; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.013 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.005 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.032 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.053 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.062 z; 100% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.02
Genuine +0.01
WER +0.00 pp
2/5 language means improve
submissive and deferential
Voice dimension · vn_STNC_reg_low
Raw target direction: ↓		-0.086 z
Mean down / no gain
Target 0.503 → 0.470
Δ score -0.033
Uncertainty and paired results

Stricter resampling range: [-0.355, +0.188] z. 2/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 2.5645 → 2.6619. Desired-direction raw change: -0.0974. CE +0.00; genuineness -0.02; WER +0.29 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.183 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.092 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.250 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.208 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.477 z; 0% paired wins; Δ WER +1.47 pp
+0.020 z
Mean up; uncertain
Target 0.503 → 0.511
Δ score +0.008
Uncertainty and paired results

Stricter resampling range: [-0.160, +0.186] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.5645 → 2.5416. Desired-direction raw change: +0.0229. CE -0.01; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.238 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.104 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.258 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.112 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.095 z; 25% paired wins; Δ WER +0.00 pp
-0.065 z
Mean down / no gain
Target 0.503 → 0.479
Δ score -0.025
Uncertainty and paired results

Stricter resampling range: [-0.246, +0.096] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.5645 → 2.6387. Desired-direction raw change: -0.0741. CE -0.02; genuineness -0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.004 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.043 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.344 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.147 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.167 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.01
Genuine +0.02
WER +0.00 pp
3/5 language means improve
dominant and assertive
Voice dimension · vn_STNC_reg_high
Raw target direction: ↑		-0.030 z
Mean down / no gain
Target 0.842 → 0.839
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.156, +0.111] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.7892 → 3.7551. Desired-direction raw change: -0.0340. CE -0.01; genuineness +0.01; WER +0.07 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.011 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.094 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.226 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.012 z; 50% paired wins; Δ WER -1.32 pp
  • ES: +0.193 z; 75% paired wins; Δ WER +1.67 pp
-0.039 z
Mean down / no gain
Target 0.842 → 0.831
Δ score -0.011
Uncertainty and paired results

Stricter resampling range: [-0.187, +0.060] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 3.7892 → 3.7450. Desired-direction raw change: -0.0441. CE +0.00; genuineness +0.01; WER +0.35 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.029 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.025 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.294 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.088 z; 50% paired wins; Δ WER +0.07 pp
  • ES: +0.009 z; 75% paired wins; Δ WER +1.67 pp
+0.019 z
Mean up; uncertain
Target 0.842 → 0.846
Δ score +0.004
Uncertainty and paired results

Stricter resampling range: [-0.160, +0.197] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 3.7892 → 3.8106. Desired-direction raw change: +0.0215. CE +0.00; genuineness -0.01; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.142 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.191 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.208 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.033 z; 50% paired wins; Δ WER -1.32 pp
  • ES: +0.219 z; 75% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.00
Genuine -0.01
WER -0.26 pp
3/5 language means improve
Sadness
Emotion · emo_Sadness
Raw target direction: ↑		+0.013 z
Mean up; uncertain
Target 0.331 → 0.336
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.050, +0.068] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1594 → 0.1691. Desired-direction raw change: +0.0096. CE +0.02; genuineness -0.02; WER +0.87 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.085 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.016 z; 50% paired wins; Δ WER +2.88 pp
  • FR: +0.091 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.011 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.052 z; 75% paired wins; Δ WER +1.47 pp
+0.017 z
Mean up; uncertain
Target 0.331 → 0.337
Δ score +0.006
Uncertainty and paired results

Stricter resampling range: [-0.030, +0.064] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1594 → 0.1728. Desired-direction raw change: +0.0134. CE -0.01; genuineness -0.06; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.008 z; 25% paired wins; Δ WER +1.39 pp
  • DE: +0.029 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.092 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.058 z; 0% paired wins; Δ WER +0.00 pp
  • ES: +0.016 z; 75% paired wins; Δ WER +0.00 pp
-0.021 z
Mean down / no gain
Target 0.331 → 0.324
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.102, +0.067] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1594 → 0.1435. Desired-direction raw change: -0.0159. CE +0.00; genuineness -0.12; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.130 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.031 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.116 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.068 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.052 z; 25% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.01
Genuine -0.06
WER +0.28 pp
4/5 language means improve
Fear
Emotion · emo_Fear
Raw target direction: ↑		+0.017 z
Mean up; uncertain
Target 0.324 → 0.330
Δ score +0.006
Uncertainty and paired results

Stricter resampling range: [-0.009, +0.048] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0420 → 0.0522. Desired-direction raw change: +0.0102. CE +0.03; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.006 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.009 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.034 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.062 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.009 z; 50% paired wins; Δ WER +0.00 pp
+0.003 z
Mean up; uncertain
Target 0.324 → 0.325
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.024, +0.036] z. 3/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0420 → 0.0439. Desired-direction raw change: +0.0019. CE +0.01; genuineness -0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.004 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.022 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.006 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.055 z; 100% paired wins; Δ WER +0.00 pp
  • ES: -0.028 z; 25% paired wins; Δ WER +0.00 pp
-0.007 z
Mean down / no gain
Target 0.324 → 0.322
Δ score -0.002
Uncertainty and paired results

Stricter resampling range: [-0.017, +0.009] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0420 → 0.0379. Desired-direction raw change: -0.0041. CE -0.01; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.012 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.010 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.014 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.022 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.020 z; 50% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.03
Genuine +0.01
WER +0.00 pp
3/5 language means improve
Disappointment
Emotion · emo_Disappointment
Raw target direction: ↑		-0.009 z
Mean down / no gain
Target 0.313 → 0.309
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.036, +0.021] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1456 → 0.1392. Desired-direction raw change: -0.0064. CE +0.04; genuineness -0.03; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.036 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.038 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.001 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.036 z; 50% paired wins; Δ WER -1.39 pp
  • ES: -0.009 z; 50% paired wins; Δ WER +0.00 pp
+0.007 z
Mean up; uncertain
Target 0.313 → 0.315
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.005, +0.020] z. 4/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1456 → 0.1508. Desired-direction raw change: +0.0052. CE -0.01; genuineness -0.07; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.006 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.027 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.009 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.013 z; 25% paired wins; Δ WER -1.39 pp
  • ES: +0.008 z; 50% paired wins; Δ WER +0.00 pp
+0.017 z
Mean up; uncertain
Target 0.313 → 0.319
Δ score +0.006
Uncertainty and paired results

Stricter resampling range: [-0.005, +0.040] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.1456 → 0.1572. Desired-direction raw change: +0.0116. CE -0.01; genuineness -0.03; WER +0.03 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.007 z; 50% paired wins; Δ WER +1.56 pp
  • DE: -0.006 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.038 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.049 z; 50% paired wins; Δ WER -1.39 pp
  • ES: -0.005 z; 50% paired wins; Δ WER +0.00 pp
Try 1.0
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.01
Genuine -0.03
WER +0.03 pp
3/5 language means improve
Infatuation
Emotion · emo_Infatuation
Raw target direction: ↑		+0.013 z
Mean up; uncertain
Target 0.376 → 0.381
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.003, +0.029] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.0265 → 0.0338. Desired-direction raw change: +0.0073. CE -0.02; genuineness -0.08; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.035 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.001 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.017 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.023 z; 75% paired wins; Δ WER +1.32 pp
  • ES: -0.007 z; 25% paired wins; Δ WER +0.00 pp
+0.012 z
Mean up; uncertain
Target 0.376 → 0.380
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.012, +0.044] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.0265 → 0.0332. Desired-direction raw change: +0.0068. CE -0.01; genuineness -0.05; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.043 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.048 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.010 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.005 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.014 z; 25% paired wins; Δ WER +0.00 pp
+0.003 z
Mean up; uncertain
Target 0.376 → 0.377
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.009, +0.015] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0265 → 0.0281. Desired-direction raw change: +0.0016. CE -0.02; genuineness -0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.015 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.022 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.002 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.007 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.001 z; 50% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.02
Genuine -0.08
WER +0.26 pp
3/5 language means improve
Pain
Emotion · emo_Pain
Raw target direction: ↑		+0.002 z
Mean up; uncertain
Target 0.369 → 0.370
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.016, +0.030] z. 2/5 language means improve. 70% of 20 paired recordings improve.

Raw head mean: 0.0643 → 0.0657. Desired-direction raw change: +0.0013. CE -0.02; genuineness -0.05; WER -0.59 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.015 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.013 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.014 z; 100% paired wins; Δ WER +0.00 pp
  • NL: +0.042 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.017 z; 50% paired wins; Δ WER -2.94 pp
+0.012 z
Mean up; uncertain
Target 0.369 → 0.373
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.010, +0.032] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0643 → 0.0712. Desired-direction raw change: +0.0069. CE +0.01; genuineness -0.00; WER +0.20 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.022 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.017 z; 25% paired wins; Δ WER +2.63 pp
  • FR: +0.039 z; 100% paired wins; Δ WER +1.32 pp
  • NL: +0.021 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.005 z; 50% paired wins; Δ WER -2.94 pp
-0.005 z
Mean down / no gain
Target 0.369 → 0.367
Δ score -0.002
Uncertainty and paired results

Stricter resampling range: [-0.027, +0.015] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0643 → 0.0617. Desired-direction raw change: -0.0026. CE -0.02; genuineness -0.10; WER -0.59 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.002 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.026 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.018 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.013 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.030 z; 25% paired wins; Δ WER -2.94 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine -0.00
WER +0.20 pp
3/5 language means improve
Intoxication Altered States of Consciousness
Emotion · emo_Intoxication_Altered_States_of_Consciousness
Raw target direction: ↑		-0.010 z
Mean down / no gain
Target 0.397 → 0.393
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.028, +0.002] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0274 → 0.0220. Desired-direction raw change: -0.0054. CE +0.03; genuineness +0.03; WER +0.84 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.002 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.039 z; 25% paired wins; Δ WER +1.56 pp
  • FR: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.003 z; 50% paired wins; Δ WER +2.63 pp
  • ES: -0.011 z; 25% paired wins; Δ WER +0.00 pp
+0.012 z
Mean up; uncertain
Target 0.397 → 0.402
Δ score +0.005
Uncertainty and paired results

Stricter resampling range: [-0.016, +0.044] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.0274 → 0.0339. Desired-direction raw change: +0.0065. CE +0.01; genuineness +0.10; WER +0.58 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.025 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.023 z; 0% paired wins; Δ WER +1.56 pp
  • FR: +0.007 z; 75% paired wins; Δ WER +1.32 pp
  • NL: +0.061 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.011 z; 50% paired wins; Δ WER +0.00 pp
-0.001 z
Mean down / no gain
Target 0.397 → 0.397
Δ score -0.000
Uncertainty and paired results

Stricter resampling range: [-0.010, +0.008] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0274 → 0.0268. Desired-direction raw change: -0.0006. CE -0.03; genuineness +0.05; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.013 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.002 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.015 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.001 z; 75% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine +0.10
WER +0.58 pp
3/5 language means improve
Contempt
Emotion · emo_Contempt
Raw target direction: ↑		+0.009 z
Mean up; uncertain
Target 0.353 → 0.356
Δ score +0.003
Uncertainty and paired results

Stricter resampling range: [-0.027, +0.044] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.0373 → 0.0413. Desired-direction raw change: +0.0040. CE +0.01; genuineness -0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.025 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.009 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.046 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.014 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.061 z; 100% paired wins; Δ WER +0.00 pp
-0.013 z
Mean down / no gain
Target 0.353 → 0.348
Δ score -0.005
Uncertainty and paired results

Stricter resampling range: [-0.030, +0.002] z. 1/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.0373 → 0.0316. Desired-direction raw change: -0.0057. CE -0.01; genuineness +0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.020 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.008 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.039 z; 0% paired wins; Δ WER +0.00 pp
  • ES: +0.007 z; 75% paired wins; Δ WER +0.00 pp
+0.002 z
Mean up; uncertain
Target 0.353 → 0.354
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.028, +0.023] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0373 → 0.0381. Desired-direction raw change: +0.0007. CE +0.03; genuineness -0.09; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.010 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.013 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.042 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.026 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.021 z; 75% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine -0.02
WER +0.00 pp
3/5 language means improve
Shame
Emotion · emo_Shame
Raw target direction: ↑		+0.009 z
Mean up; uncertain
Target 0.356 → 0.359
Δ score +0.003
Uncertainty and paired results

Stricter resampling range: [-0.011, +0.033] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0168 → 0.0214. Desired-direction raw change: +0.0045. CE -0.01; genuineness -0.05; WER -0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.001 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.021 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.045 z; 100% paired wins; Δ WER +0.00 pp
  • NL: -0.017 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.008 z; 25% paired wins; Δ WER -1.67 pp
+0.005 z
Mean up; uncertain
Target 0.356 → 0.357
Δ score +0.002
Uncertainty and paired results

Stricter resampling range: [-0.012, +0.025] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.0168 → 0.0196. Desired-direction raw change: +0.0027. CE +0.02; genuineness -0.01; WER +0.19 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.001 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.018 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.034 z; 100% paired wins; Δ WER +2.63 pp
  • NL: -0.014 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.011 z; 25% paired wins; Δ WER -1.67 pp
+0.000 z
Mean up; uncertain
Target 0.356 → 0.356
Δ score +0.000
Uncertainty and paired results

Stricter resampling range: [-0.021, +0.026] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.0168 → 0.0171. Desired-direction raw change: +0.0002. CE -0.02; genuineness +0.01; WER -0.06 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.015 z; 0% paired wins; Δ WER +0.00 pp
  • DE: +0.025 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.028 z; 100% paired wins; Δ WER +1.39 pp
  • NL: -0.028 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.009 z; 50% paired wins; Δ WER -1.67 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE -0.01
Genuine -0.05
WER -0.33 pp
3/5 language means improve
Strong Playful Style
Talking style · vn_S_PLAY_reg
Raw target direction: ↑		+0.007 z
Mean up; uncertain
Target 0.344 → 0.348
Δ score +0.004
Uncertainty and paired results

Stricter resampling range: [-0.160, +0.161] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 1.7153 → 1.7243. Desired-direction raw change: +0.0090. CE +0.01; genuineness +0.03; WER +0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.003 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.007 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.252 z; 75% paired wins; Δ WER +1.32 pp
  • NL: +0.050 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.262 z; 25% paired wins; Δ WER +0.00 pp
+0.007 z
Mean up; uncertain
Target 0.344 → 0.343
Δ score -0.001

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.256, +0.212] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 1.7153 → 1.7237. Desired-direction raw change: +0.0085. CE -0.01; genuineness +0.03; WER +0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.090 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.176 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.266 z; 75% paired wins; Δ WER +0.00 pp
  • NL: +0.066 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.386 z; 25% paired wins; Δ WER +1.67 pp
-0.048 z
Mean down / no gain
Target 0.344 → 0.328
Δ score -0.016
Uncertainty and paired results

Stricter resampling range: [-0.138, +0.014] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 1.7153 → 1.6529. Desired-direction raw change: -0.0623. CE -0.05; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.033 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.021 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.009 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.038 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.204 z; 50% paired wins; Δ WER +0.00 pp
Try 0.5
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.01
Genuine +0.03
WER +0.26 pp
3/5 language means improve
Very informal/casual, not formal
Style contrast · casual_not_formal
Raw target direction: ↑		-0.033 z
Mean down / no gain
Target 0.220 → 0.213
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.206, +0.213] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.6237 → 2.5817. Desired-direction raw change: -0.0420. CE +0.02; genuineness +0.00; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.371 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.025 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.112 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.260 z; 25% paired wins; Δ WER -1.32 pp
  • ES: -0.139 z; 25% paired wins; Δ WER +0.00 pp

Opposite head vn_S_FORM_reg: desired-direction change -0.091 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

-0.084 z
Mean down / no gain
Target 0.220 → 0.196
Δ score -0.024
Uncertainty and paired results

Stricter resampling range: [-0.240, +0.124] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.6237 → 2.5169. Desired-direction raw change: -0.1068. CE +0.02; genuineness -0.05; WER +0.01 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.217 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.174 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.023 z; 75% paired wins; Δ WER +1.39 pp
  • NL: -0.239 z; 0% paired wins; Δ WER -1.32 pp
  • ES: -0.246 z; 50% paired wins; Δ WER +0.00 pp

Opposite head vn_S_FORM_reg: desired-direction change -0.136 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

+0.005 z
Mean up; uncertain
Target 0.220 → 0.217
Δ score -0.003

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.252, +0.187] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.6237 → 2.6297. Desired-direction raw change: +0.0061. CE +0.04; genuineness +0.00; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.224 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.178 z; 100% paired wins; Δ WER +0.00 pp
  • FR: +0.033 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.440 z; 25% paired wins; Δ WER -1.32 pp
  • ES: +0.029 z; 50% paired wins; Δ WER +0.00 pp

Opposite head vn_S_FORM_reg: desired-direction change -0.074 z (lower raw is desired). Both heads must improve for a clearer style-contrast result.

Try 1.0
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.
Primary and opposite heads reported in cell details.		CE +0.04
Genuine +0.00
WER -0.26 pp
4/5 language means improve
very fast speaking tempo
Voice dimension · vn_TEMP_reg_high
Raw target direction: ↑		-0.014 z
Mean down / no gain
Target 0.588 → 0.583
Δ score -0.005
Uncertainty and paired results

Stricter resampling range: [-0.136, +0.072] z. 3/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 2.3865 → 2.3693. Desired-direction raw change: -0.0172. CE +0.03; genuineness +0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.073 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.072 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.209 z; 0% paired wins; Δ WER +0.00 pp
  • NL: +0.020 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.028 z; 25% paired wins; Δ WER +0.00 pp
+0.003 z
Mean up; uncertain
Target 0.588 → 0.590
Δ score +0.001
Uncertainty and paired results

Stricter resampling range: [-0.048, +0.070] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 2.3865 → 2.3906. Desired-direction raw change: +0.0040. CE +0.03; genuineness +0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.101 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.035 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.053 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.049 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.045 z; 25% paired wins; Δ WER +0.00 pp
-0.006 z
Mean down / no gain
Target 0.588 → 0.586
Δ score -0.002
Uncertainty and paired results

Stricter resampling range: [-0.051, +0.040] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.3865 → 2.3796. Desired-direction raw change: -0.0069. CE +0.03; genuineness +0.03; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.069 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.004 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.002 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.081 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.010 z; 50% paired wins; Δ WER +0.00 pp
Try 0.75
Positive raw mean; uncertain
Very small scorer change; audible benefit unproven.		CE +0.03
Genuine +0.00
WER +0.00 pp
2/5 language means improve
Anger
Emotion · emo_Anger
Raw target direction: ↑		-0.012 z
Mean down / no gain
Target 0.306 → 0.302
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.031, +0.016] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0648 → 0.0577. Desired-direction raw change: -0.0071. CE +0.01; genuineness -0.01; WER +0.83 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.025 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.004 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.029 z; 75% paired wins; Δ WER +4.17 pp
  • NL: -0.026 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.035 z; 25% paired wins; Δ WER +0.00 pp
-0.011 z
Mean down / no gain
Target 0.306 → 0.303
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.024, +0.002] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0648 → 0.0586. Desired-direction raw change: -0.0062. CE +0.01; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.008 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.009 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.024 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.023 z; 25% paired wins; Δ WER +0.00 pp
-0.000 z
Mean down / no gain
Target 0.306 → 0.306
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.022, +0.022] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0648 → 0.0646. Desired-direction raw change: -0.0002. CE +0.00; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.009 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.020 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.005 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.028 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.033 z; 75% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Embarrassment
Emotion · emo_Embarrassment
Raw target direction: ↑		-0.004 z
Mean down / no gain
Target 0.304 → 0.302
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.017, +0.010] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0332 → 0.0316. Desired-direction raw change: -0.0016. CE -0.01; genuineness -0.07; WER +0.25 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.005 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.018 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.023 z; 50% paired wins; Δ WER -1.39 pp
  • NL: -0.010 z; 25% paired wins; Δ WER +2.63 pp
  • ES: -0.008 z; 25% paired wins; Δ WER +0.00 pp
-0.010 z
Mean down / no gain
Target 0.304 → 0.300
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.022, +0.003] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.0332 → 0.0289. Desired-direction raw change: -0.0043. CE -0.02; genuineness -0.07; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.003 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.002 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.009 z; 50% paired wins; Δ WER -1.39 pp
  • NL: -0.025 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.021 z; 25% paired wins; Δ WER +0.00 pp
-0.001 z
Mean down / no gain
Target 0.304 → 0.303
Δ score -0.000
Uncertainty and paired results

Stricter resampling range: [-0.008, +0.005] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0332 → 0.0328. Desired-direction raw change: -0.0003. CE -0.01; genuineness -0.01; WER -0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.006 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.007 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.001 z; 75% paired wins; Δ WER -1.39 pp
  • NL: -0.011 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.004 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Bitterness
Emotion · emo_Bitterness
Raw target direction: ↑		-0.035 z
Mean down / no gain
Target 0.353 → 0.340
Δ score -0.013
Uncertainty and paired results

Stricter resampling range: [-0.078, -0.010] z. 0/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.0641 → 0.0498. Desired-direction raw change: -0.0143. CE +0.03; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.014 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.016 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.112 z; 0% paired wins; Δ WER +0.00 pp
  • NL: -0.026 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.007 z; 25% paired wins; Δ WER +0.00 pp
-0.003 z
Mean down / no gain
Target 0.353 → 0.352
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.088, +0.112] z. 2/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 0.0641 → 0.0630. Desired-direction raw change: -0.0011. CE +0.01; genuineness -0.07; WER +0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.019 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.175 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.118 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.069 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.021 z; 50% paired wins; Δ WER +1.67 pp
-0.022 z
Mean down / no gain
Target 0.353 → 0.345
Δ score -0.008
Uncertainty and paired results

Stricter resampling range: [-0.075, +0.017] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.0641 → 0.0551. Desired-direction raw change: -0.0090. CE +0.02; genuineness -0.11; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.023 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.016 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.108 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.025 z; 50% paired wins; Δ WER +1.39 pp
  • ES: -0.016 z; 25% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Jealousy and Envy
Emotion · emo_Jealousy_and_Envy
Raw target direction: ↑		-0.009 z
Mean down / no gain
Target 0.318 → 0.315
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.012, -0.006] z. 0/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.0033 → -0.0007. Desired-direction raw change: -0.0040. CE +0.03; genuineness -0.06; WER -0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.009 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.011 z; 25% paired wins; Δ WER -1.56 pp
  • FR: -0.008 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.004 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.012 z; 25% paired wins; Δ WER +0.00 pp
-0.004 z
Mean down / no gain
Target 0.318 → 0.316
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.014, +0.007] z. 2/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.0033 → 0.0014. Desired-direction raw change: -0.0018. CE +0.03; genuineness +0.05; WER +1.35 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.007 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.008 z; 50% paired wins; Δ WER -1.56 pp
  • FR: -0.013 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.007 z; 25% paired wins; Δ WER +8.33 pp
  • ES: -0.016 z; 0% paired wins; Δ WER +0.00 pp
-0.010 z
Mean down / no gain
Target 0.318 → 0.314
Δ score -0.004
Uncertainty and paired results

Stricter resampling range: [-0.022, +0.004] z. 1/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.0033 → -0.0012. Desired-direction raw change: -0.0044. CE +0.04; genuineness +0.04; WER -0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.011 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.021 z; 25% paired wins; Δ WER -1.56 pp
  • FR: -0.007 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.006 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.026 z; 25% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Strong Narrator Style
Talking style · vn_S_NARR_reg
Raw target direction: ↑		-0.010 z
Mean down / no gain
Target 0.941 → 0.940
Δ score -0.000
Uncertainty and paired results

Stricter resampling range: [-0.161, +0.147] z. 3/5 language means improve. 60% of 20 paired recordings improve.

Raw head mean: 3.7853 → 3.7710. Desired-direction raw change: -0.0143. CE -0.03; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.225 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.055 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.008 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.155 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.182 z; 25% paired wins; Δ WER +0.00 pp
-0.066 z
Mean down / no gain
Target 0.941 → 0.935
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.188, +0.069] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 3.7853 → 3.6900. Desired-direction raw change: -0.0953. CE -0.05; genuineness -0.04; WER +0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.141 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.041 z; 25% paired wins; Δ WER +1.56 pp
  • FR: -0.035 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.252 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.145 z; 25% paired wins; Δ WER +0.00 pp
-0.017 z
Mean down / no gain
Target 0.941 → 0.934
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.074, +0.057] z. 1/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 3.7853 → 3.7606. Desired-direction raw change: -0.0247. CE -0.02; genuineness +0.00; WER +0.28 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.070 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.018 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.105 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.088 z; 25% paired wins; Δ WER +1.39 pp
  • ES: -0.016 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Strong Dramatic Style
Talking style · vn_S_DRAM_reg
Raw target direction: ↑		-0.011 z
Mean down / no gain
Target 0.496 → 0.494
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.100, +0.071] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 2.8025 → 2.7896. Desired-direction raw change: -0.0129. CE -0.02; genuineness -0.01; WER -1.32 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.010 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.032 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.111 z; 75% paired wins; Δ WER -3.95 pp
  • NL: -0.151 z; 50% paired wins; Δ WER -2.63 pp
  • ES: -0.038 z; 25% paired wins; Δ WER +0.00 pp
-0.181 z
Mean down / no gain
Target 0.496 → 0.429
Δ score -0.067
Uncertainty and paired results

Stricter resampling range: [-0.292, -0.055] z. 1/5 language means improve. 20% of 20 paired recordings improve.

Raw head mean: 2.8025 → 2.5967. Desired-direction raw change: -0.2058. CE -0.02; genuineness -0.05; WER -1.05 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.139 z; 0% paired wins; Δ WER +0.00 pp
  • DE: -0.266 z; 0% paired wins; Δ WER +1.32 pp
  • FR: +0.014 z; 50% paired wins; Δ WER -3.95 pp
  • NL: -0.181 z; 50% paired wins; Δ WER -2.63 pp
  • ES: -0.330 z; 0% paired wins; Δ WER +0.00 pp
-0.057 z
Mean down / no gain
Target 0.496 → 0.476
Δ score -0.020
Uncertainty and paired results

Stricter resampling range: [-0.181, +0.054] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 2.8025 → 2.7371. Desired-direction raw change: -0.0654. CE -0.02; genuineness -0.00; WER -1.32 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.018 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.119 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.094 z; 50% paired wins; Δ WER -3.95 pp
  • NL: -0.031 z; 50% paired wins; Δ WER -2.63 pp
  • ES: -0.249 z; 0% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
emotionally detached and unexposed
Voice dimension · vn_VULN_reg_low
Raw target direction: ↓		-0.232 z
Mean down / no gain
Target 0.711 → 0.632
Δ score -0.079
Uncertainty and paired results

Stricter resampling range: [-0.293, -0.155] z. 0/5 language means improve. 25% of 20 paired recordings improve.

Raw head mean: 2.1107 → 2.3409. Desired-direction raw change: -0.2301. CE -0.01; genuineness -0.00; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.297 z; 0% paired wins; Δ WER +0.00 pp
  • DE: -0.277 z; 25% paired wins; Δ WER -1.32 pp
  • FR: -0.139 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.159 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.290 z; 25% paired wins; Δ WER +0.00 pp
-0.146 z
Mean down / no gain
Target 0.711 → 0.662
Δ score -0.048
Uncertainty and paired results

Stricter resampling range: [-0.355, +0.044] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.1107 → 2.2550. Desired-direction raw change: -0.1443. CE -0.04; genuineness -0.07; WER +0.03 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.100 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.006 z; 50% paired wins; Δ WER -1.32 pp
  • FR: -0.432 z; 0% paired wins; Δ WER +0.00 pp
  • NL: -0.303 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.098 z; 50% paired wins; Δ WER +1.47 pp
-0.012 z
Mean down / no gain
Target 0.711 → 0.708
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.096, +0.065] z. 2/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.1107 → 2.1231. Desired-direction raw change: -0.0124. CE +0.01; genuineness -0.10; WER +0.86 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.009 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.027 z; 50% paired wins; Δ WER -1.32 pp
  • FR: +0.107 z; 75% paired wins; Δ WER +2.63 pp
  • NL: +0.014 z; 25% paired wins; Δ WER +1.32 pp
  • ES: -0.147 z; 25% paired wins; Δ WER +1.67 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Impatience and Irritability
Emotion · emo_Impatience_and_Irritability
Raw target direction: ↑		-0.015 z
Mean down / no gain
Target 0.316 → 0.309
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.144, +0.089] z. 3/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1645 → 0.1544. Desired-direction raw change: -0.0101. CE -0.02; genuineness -0.04; WER +1.16 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.070 z; 100% paired wins; Δ WER +0.00 pp
  • DE: +0.116 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.052 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.209 z; 0% paired wins; Δ WER +1.39 pp
  • ES: +0.001 z; 75% paired wins; Δ WER +4.41 pp
-0.033 z
Mean down / no gain
Target 0.316 → 0.303
Δ score -0.013
Uncertainty and paired results

Stricter resampling range: [-0.136, +0.029] z. 3/5 language means improve. 65% of 20 paired recordings improve.

Raw head mean: 0.1645 → 0.1423. Desired-direction raw change: -0.0222. CE +0.02; genuineness +0.12; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.035 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.036 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.015 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.202 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.025 z; 75% paired wins; Δ WER +0.00 pp
-0.025 z
Mean down / no gain
Target 0.316 → 0.305
Δ score -0.011
Uncertainty and paired results

Stricter resampling range: [-0.206, +0.089] z. 4/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1645 → 0.1476. Desired-direction raw change: -0.0170. CE -0.01; genuineness +0.02; WER +1.67 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.014 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.115 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.052 z; 50% paired wins; Δ WER +8.33 pp
  • NL: -0.353 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.048 z; 75% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Strong teacher-like, didactic delivery
Talking style · vn_S_TECH_reg
Raw target direction: ↑		-0.017 z
Mean down / no gain
Target 0.931 → 0.925
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.136, +0.083] z. 3/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 2.3946 → 2.3737. Desired-direction raw change: -0.0209. CE +0.01; genuineness -0.03; WER +1.84 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.178 z; 0% paired wins; Δ WER +0.00 pp
  • DE: -0.125 z; 0% paired wins; Δ WER +0.00 pp
  • FR: +0.068 z; 75% paired wins; Δ WER +9.21 pp
  • NL: +0.100 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.049 z; 25% paired wins; Δ WER +0.00 pp
-0.034 z
Mean down / no gain
Target 0.931 → 0.920
Δ score -0.011
Uncertainty and paired results

Stricter resampling range: [-0.111, +0.036] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 2.3946 → 2.3534. Desired-direction raw change: -0.0412. CE +0.01; genuineness -0.00; WER +1.15 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.029 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.113 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.110 z; 25% paired wins; Δ WER +1.32 pp
  • NL: +0.053 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.031 z; 50% paired wins; Δ WER +4.41 pp
-0.024 z
Mean down / no gain
Target 0.931 → 0.926
Δ score -0.005
Uncertainty and paired results

Stricter resampling range: [-0.137, +0.105] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 2.3946 → 2.3651. Desired-direction raw change: -0.0295. CE -0.00; genuineness -0.02; WER +2.78 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.197 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.023 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.039 z; 50% paired wins; Δ WER +13.89 pp
  • NL: +0.191 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.053 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Amusement
Emotion · emo_Amusement
Raw target direction: ↑		-0.104 z
Mean down / no gain
Target 0.583 → 0.549
Δ score -0.034
Uncertainty and paired results

Stricter resampling range: [-0.222, +0.005] z. 1/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.6843 → 0.5936. Desired-direction raw change: -0.0907. CE +0.03; genuineness -0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.129 z; 25% paired wins; Δ WER +0.00 pp
  • DE: -0.141 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.010 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.038 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.281 z; 25% paired wins; Δ WER +0.00 pp
-0.077 z
Mean down / no gain
Target 0.583 → 0.555
Δ score -0.028
Uncertainty and paired results

Stricter resampling range: [-0.175, +0.042] z. 1/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.6843 → 0.6173. Desired-direction raw change: -0.0670. CE -0.01; genuineness -0.01; WER +0.23 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.060 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.103 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.114 z; 0% paired wins; Δ WER +2.63 pp
  • NL: +0.110 z; 75% paired wins; Δ WER +0.00 pp
  • ES: -0.219 z; 25% paired wins; Δ WER -1.47 pp
-0.018 z
Mean down / no gain
Target 0.583 → 0.573
Δ score -0.010
Uncertainty and paired results

Stricter resampling range: [-0.154, +0.095] z. 2/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.6843 → 0.6690. Desired-direction raw change: -0.0153. CE +0.01; genuineness +0.02; WER -0.29 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.145 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.228 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.045 z; 25% paired wins; Δ WER +0.00 pp
  • NL: +0.048 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.007 z; 25% paired wins; Δ WER -1.47 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Triumph
Emotion · emo_Triumph
Raw target direction: ↑		-0.065 z
Mean down / no gain
Target 0.479 → 0.474
Δ score -0.006
Uncertainty and paired results

Stricter resampling range: [-0.184, +0.009] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.2444 → 0.2103. Desired-direction raw change: -0.0341. CE -0.04; genuineness +0.03; WER -0.26 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.007 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.280 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.015 z; 75% paired wins; Δ WER -1.32 pp
  • NL: -0.040 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.026 z; 50% paired wins; Δ WER +0.00 pp
-0.052 z
Mean down / no gain
Target 0.479 → 0.486
Δ score +0.007

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.160, +0.023] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.2444 → 0.2170. Desired-direction raw change: -0.0274. CE +0.01; genuineness +0.01; WER -0.53 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.001 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.241 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.027 z; 75% paired wins; Δ WER -2.63 pp
  • NL: -0.039 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.046 z; 50% paired wins; Δ WER +0.00 pp
-0.104 z
Mean down / no gain
Target 0.479 → 0.476
Δ score -0.003
Uncertainty and paired results

Stricter resampling range: [-0.380, +0.084] z. 2/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.2444 → 0.1896. Desired-direction raw change: -0.0547. CE -0.05; genuineness +0.03; WER -0.53 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.062 z; 0% paired wins; Δ WER +0.00 pp
  • DE: -0.591 z; 0% paired wins; Δ WER +0.00 pp
  • FR: +0.143 z; 50% paired wins; Δ WER -2.63 pp
  • NL: -0.018 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.010 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Sexual Lust
Emotion · emo_Sexual_Lust
Raw target direction: ↑		-0.055 z
Mean down / no gain
Target 0.357 → 0.339
Δ score -0.018
Uncertainty and paired results

Stricter resampling range: [-0.133, +0.011] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.0732 → 0.0459. Desired-direction raw change: -0.0273. CE -0.00; genuineness +0.00; WER +0.29 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.165 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.020 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.113 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.024 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.008 z; 50% paired wins; Δ WER +1.47 pp
-0.124 z
Mean down / no gain
Target 0.357 → 0.312
Δ score -0.044
Uncertainty and paired results

Stricter resampling range: [-0.375, +0.028] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.0732 → 0.0113. Desired-direction raw change: -0.0619. CE +0.00; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.168 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.030 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.514 z; 0% paired wins; Δ WER +0.00 pp
  • NL: +0.002 z; 75% paired wins; Δ WER +0.00 pp
  • ES: +0.028 z; 50% paired wins; Δ WER +0.00 pp
-0.110 z
Mean down / no gain
Target 0.357 → 0.318
Δ score -0.038
Uncertainty and paired results

Stricter resampling range: [-0.284, +0.003] z. 1/5 language means improve. 25% of 20 paired recordings improve.

Raw head mean: 0.0732 → 0.0184. Desired-direction raw change: -0.0548. CE -0.04; genuineness -0.00; WER -0.07 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.186 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.011 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.358 z; 0% paired wins; Δ WER +1.32 pp
  • NL: -0.012 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.006 z; 25% paired wins; Δ WER -1.67 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Contemplation
Emotion · emo_Contemplation
Raw target direction: ↑		-0.079 z
Mean down / no gain
Target 0.940 → 0.948
Δ score +0.008

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.353, +0.129] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 1.8347 → 1.7879. Desired-direction raw change: -0.0468. CE +0.00; genuineness -0.04; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.104 z; 75% paired wins; Δ WER +0.00 pp
  • DE: +0.147 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.406 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.340 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.098 z; 25% paired wins; Δ WER +0.00 pp
-0.066 z
Mean down / no gain
Target 0.940 → 0.931
Δ score -0.009
Uncertainty and paired results

Stricter resampling range: [-0.418, +0.287] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 1.8347 → 1.7958. Desired-direction raw change: -0.0389. CE +0.01; genuineness -0.06; WER +0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.385 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.104 z; 50% paired wins; Δ WER +1.56 pp
  • FR: -0.557 z; 0% paired wins; Δ WER +0.00 pp
  • NL: -0.314 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.260 z; 75% paired wins; Δ WER +0.00 pp
-0.093 z
Mean down / no gain
Target 0.940 → 0.938
Δ score -0.002
Uncertainty and paired results

Stricter resampling range: [-0.504, +0.284] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 1.8347 → 1.7797. Desired-direction raw change: -0.0550. CE +0.02; genuineness -0.08; WER +0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.370 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.253 z; 100% paired wins; Δ WER +1.56 pp
  • FR: -0.316 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.715 z; 0% paired wins; Δ WER +0.00 pp
  • ES: -0.059 z; 25% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Strong Newsreader Style
Talking style · vn_S_NEWS_reg
Raw target direction: ↑		-0.108 z
Mean down / no gain
Target 0.815 → 0.788
Δ score -0.027
Uncertainty and paired results

Stricter resampling range: [-0.425, +0.170] z. 1/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.0908 → 1.0312. Desired-direction raw change: -0.0596. CE -0.03; genuineness +0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.104 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.053 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.068 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.640 z; 0% paired wins; Δ WER +0.00 pp
  • ES: +0.323 z; 75% paired wins; Δ WER +0.00 pp
-0.069 z
Mean down / no gain
Target 0.815 → 0.815
Δ score +0.000

Mixed score views: raw and CDF-transformed means move differently; this is not agreement between both summaries.

Uncertainty and paired results

Stricter resampling range: [-0.337, +0.070] z. 4/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 1.0908 → 1.0529. Desired-direction raw change: -0.0379. CE -0.05; genuineness -0.07; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.071 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.061 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.049 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.595 z; 0% paired wins; Δ WER +0.00 pp
  • ES: +0.069 z; 50% paired wins; Δ WER +0.00 pp
-0.127 z
Mean down / no gain
Target 0.815 → 0.794
Δ score -0.020
Uncertainty and paired results

Stricter resampling range: [-0.471, +0.123] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 1.0908 → 1.0206. Desired-direction raw change: -0.0702. CE -0.02; genuineness -0.01; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.168 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.174 z; 50% paired wins; Δ WER +0.00 pp
  • FR: +0.088 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.672 z; 0% paired wins; Δ WER +0.00 pp
  • ES: -0.060 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Contentment
Emotion · emo_Contentment
Raw target direction: ↑		-0.088 z
Mean down / no gain
Target 0.982 → 0.970
Δ score -0.012
Uncertainty and paired results

Stricter resampling range: [-0.257, +0.056] z. 2/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 1.9641 → 1.9195. Desired-direction raw change: -0.0446. CE -0.05; genuineness +0.08; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.094 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.012 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.012 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.321 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.215 z; 25% paired wins; Δ WER +0.00 pp
-0.203 z
Mean down / no gain
Target 0.982 → 0.981
Δ score -0.001
Uncertainty and paired results

Stricter resampling range: [-0.603, +0.191] z. 2/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 1.9641 → 1.8618. Desired-direction raw change: -0.1023. CE -0.03; genuineness +0.11; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.094 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.336 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.215 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.445 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.785 z; 0% paired wins; Δ WER +0.00 pp
-0.191 z
Mean down / no gain
Target 0.982 → 0.973
Δ score -0.009
Uncertainty and paired results

Stricter resampling range: [-0.368, -0.025] z. 1/5 language means improve. 25% of 20 paired recordings improve.

Raw head mean: 1.9641 → 1.8677. Desired-direction raw change: -0.0964. CE -0.08; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.059 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.102 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.252 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.463 z; 0% paired wins; Δ WER +0.00 pp
  • ES: -0.199 z; 0% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Pride
Emotion · emo_Pride
Raw target direction: ↑		-0.167 z
Mean down / no gain
Target 0.773 → 0.756
Δ score -0.017
Uncertainty and paired results

Stricter resampling range: [-0.443, +0.027] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.7365 → 0.6494. Desired-direction raw change: -0.0870. CE -0.01; genuineness +0.00; WER +1.94 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.068 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.623 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.174 z; 25% paired wins; Δ WER +9.72 pp
  • NL: -0.108 z; 50% paired wins; Δ WER +0.00 pp
  • ES: +0.000 z; 50% paired wins; Δ WER +0.00 pp
-0.103 z
Mean down / no gain
Target 0.773 → 0.766
Δ score -0.007
Uncertainty and paired results

Stricter resampling range: [-0.313, +0.033] z. 2/5 language means improve. 40% of 20 paired recordings improve.

Raw head mean: 0.7365 → 0.6830. Desired-direction raw change: -0.0534. CE -0.02; genuineness +0.02; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.041 z; 50% paired wins; Δ WER +0.00 pp
  • DE: -0.481 z; 0% paired wins; Δ WER +0.00 pp
  • FR: -0.060 z; 50% paired wins; Δ WER +0.00 pp
  • NL: +0.031 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.045 z; 50% paired wins; Δ WER +0.00 pp
-0.192 z
Mean down / no gain
Target 0.773 → 0.734
Δ score -0.039
Uncertainty and paired results

Stricter resampling range: [-0.317, -0.063] z. 0/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.7365 → 0.6368. Desired-direction raw change: -0.0996. CE -0.02; genuineness +0.02; WER +0.33 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.002 z; 75% paired wins; Δ WER +0.00 pp
  • DE: -0.373 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.273 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.118 z; 25% paired wins; Δ WER +0.00 pp
  • ES: -0.192 z; 25% paired wins; Δ WER +1.67 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Confusion
Emotion · emo_Confusion
Raw target direction: ↑		-0.116 z
Mean down / no gain
Target 0.369 → 0.343
Δ score -0.026
Uncertainty and paired results

Stricter resampling range: [-0.448, +0.092] z. 2/5 language means improve. 55% of 20 paired recordings improve.

Raw head mean: 0.1596 → 0.0979. Desired-direction raw change: -0.0617. CE +0.02; genuineness -0.01; WER +0.53 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.048 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.046 z; 25% paired wins; Δ WER +0.00 pp
  • FR: -0.003 z; 50% paired wins; Δ WER +2.63 pp
  • NL: -0.717 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.139 z; 75% paired wins; Δ WER +0.00 pp
-0.108 z
Mean down / no gain
Target 0.369 → 0.343
Δ score -0.026
Uncertainty and paired results

Stricter resampling range: [-0.389, +0.054] z. 3/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.1596 → 0.1020. Desired-direction raw change: -0.0576. CE +0.07; genuineness +0.00; WER +0.00 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.067 z; 100% paired wins; Δ WER +0.00 pp
  • DE: -0.072 z; 0% paired wins; Δ WER +0.00 pp
  • FR: +0.041 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.601 z; 0% paired wins; Δ WER +0.00 pp
  • ES: +0.025 z; 25% paired wins; Δ WER +0.00 pp
-0.158 z
Mean down / no gain
Target 0.369 → 0.327
Δ score -0.042
Uncertainty and paired results

Stricter resampling range: [-0.496, +0.021] z. 3/5 language means improve. 50% of 20 paired recordings improve.

Raw head mean: 0.1596 → 0.0755. Desired-direction raw change: -0.0840. CE +0.03; genuineness -0.09; WER +0.31 percentage points. Frame-cap rate 0% → 0%.

  • EN: +0.024 z; 75% paired wins; Δ WER +1.56 pp
  • DE: -0.048 z; 25% paired wins; Δ WER +0.00 pp
  • FR: +0.020 z; 75% paired wins; Δ WER +0.00 pp
  • NL: -0.795 z; 25% paired wins; Δ WER +0.00 pp
  • ES: +0.012 z; 50% paired wins; Δ WER +0.00 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Pleasure Ecstasy
Emotion · emo_Pleasure_Ecstasy
Raw target direction: ↑		-0.127 z
Mean down / no gain
Target 0.486 → 0.438
Δ score -0.048
Uncertainty and paired results

Stricter resampling range: [-0.344, +0.004] z. 2/5 language means improve. 45% of 20 paired recordings improve.

Raw head mean: 0.3878 → 0.3101. Desired-direction raw change: -0.0777. CE -0.00; genuineness -0.03; WER -0.67 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.053 z; 50% paired wins; Δ WER +0.00 pp
  • DE: +0.017 z; 75% paired wins; Δ WER +0.00 pp
  • FR: +0.001 z; 50% paired wins; Δ WER +0.00 pp
  • NL: -0.086 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.515 z; 0% paired wins; Δ WER -3.33 pp
-0.141 z
Mean down / no gain
Target 0.486 → 0.432
Δ score -0.054
Uncertainty and paired results

Stricter resampling range: [-0.382, +0.012] z. 1/5 language means improve. 35% of 20 paired recordings improve.

Raw head mean: 0.3878 → 0.3020. Desired-direction raw change: -0.0859. CE +0.03; genuineness -0.01; WER -0.67 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.042 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.051 z; 75% paired wins; Δ WER +0.00 pp
  • FR: -0.049 z; 25% paired wins; Δ WER +0.00 pp
  • NL: -0.079 z; 50% paired wins; Δ WER +0.00 pp
  • ES: -0.584 z; 0% paired wins; Δ WER -3.33 pp
-0.176 z
Mean down / no gain
Target 0.486 → 0.423
Δ score -0.063
Uncertainty and paired results

Stricter resampling range: [-0.463, +0.073] z. 1/5 language means improve. 30% of 20 paired recordings improve.

Raw head mean: 0.3878 → 0.2803. Desired-direction raw change: -0.1076. CE +0.05; genuineness -0.03; WER +0.12 percentage points. Frame-cap rate 0% → 0%.

  • EN: -0.133 z; 25% paired wins; Δ WER +0.00 pp
  • DE: +0.190 z; 50% paired wins; Δ WER +0.00 pp
  • FR: -0.238 z; 0% paired wins; Δ WER +0.00 pp
  • NL: -0.074 z; 50% paired wins; Δ WER +3.95 pp
  • ES: -0.626 z; 25% paired wins; Δ WER -3.33 pp
Use the no-adapter base
No positive raw-target mean at tested scales		Base is the comparison anchor; no additional adapter is suggested by this grid.
Method, uncertainty and limits for readers and agents

There are 20 paired recordings per strength:5 languages ×2 texts ×2 seeds. One fixed reference voice is used per language. All outputs, including termination/frame-cap failures, remain included. The 0 strength baseline uses the same bucket prompt, not a neutral unconditioned prompt.

We enumerate all 3,125 bootstrap resamples of the 5 language clusters, retaining the 2 texts and 2 seeds inside each language. The displayed stricter interval is 98.33% (a conservative allowance for exploring 3 strengths within an adapter). These are exploratory uncertainty estimates with only 5 clusters. There is no global multiplicity claim across 72 adapters. The selected strength is chosen on the same grid, not an independent validation set.

The starting strength is the largest target gain among clearer strengths when any exist; otherwise it is the largest positive mean, marked uncertain. If all means are non-positive, the base is suggested. More strength is not reliably better. The normalized 0–1 mean and raw/z mean can move differently because the CDF is nonlinear; all are exported.

The judges are frozen Humaneness EARS model heads and the existing multilingual ASR scorer, applied to complete waveforms before MP 3 encoding. These are model-based scores, not human opinions. Two translated texts and one reference voice per language do not establish generalization to new voices or complex acting scenes. Language effects and reference-voice effects cannot be separated here. Evaluation also reuses the EARS model family involved in training-data screening. Listen to the matched players before adopting a setting.

Download all aggregate and per-language results (CSV) · Machine-readable results (JSON) · Full method and reproduction notes · Analysis source

Training and evidence inventory

BucketMeaningTraining / listening statusVerified training clips
emo_AffectionIntense AffectionMatched predictions published228
emo_AmusementIntense AmusementMatched predictions published211
emo_AngerIntense AngerMatched predictions published132
emo_Astonishment_SurpriseIntense Astonishment SurpriseMatched predictions published61
emo_AweIntense AweMatched predictions published113
emo_BitternessIntense BitternessMatched predictions published136
emo_ConcentrationIntense ConcentrationMatched predictions published29
emo_ConfusionIntense ConfusionMatched predictions published39
emo_ContemplationIntense ContemplationMatched predictions published179
emo_ContemptIntense ContemptMatched predictions published94
emo_ContentmentIntense ContentmentMatched predictions published81
emo_DisappointmentIntense DisappointmentMatched predictions published196
emo_DisgustIntense DisgustMatched predictions published41
emo_DistressIntense DistressMatched predictions published309
emo_DoubtIntense DoubtMatched predictions published29
emo_ElationIntense ElationMatched predictions published62
emo_EmbarrassmentIntense EmbarrassmentMatched predictions published28
emo_Emotional_NumbnessIntense Emotional NumbnessMatched predictions published29
emo_Fatigue_ExhaustionIntense Fatigue ExhaustionMatched predictions published192
emo_FearIntense FearMatched predictions published240
emo_HelplessnessIntense HelplessnessMatched predictions published324
emo_Hope_Enthusiasm_OptimismIntense Hope Enthusiasm OptimismMatched predictions published96
emo_Impatience_and_IrritabilityIntense Impatience and IrritabilityMatched predictions published157
emo_InfatuationIntense InfatuationMatched predictions published31
emo_InterestIntense InterestMatched predictions published140
emo_Intoxication_Altered_States_of_ConsciousnessIntense Intoxication Altered States of ConsciousnessMatched predictions published31
emo_Jealousy_and_EnvyIntense Jealousy and EnvyMatched predictions published29
emo_LongingIntense LongingMatched predictions published113
emo_Malevolence_MaliceIntense Malevolence MaliceMatched predictions published29
emo_PainIntense PainMatched predictions published187
emo_Pleasure_EcstasyIntense Pleasure EcstasyMatched predictions published28
emo_PrideIntense PrideMatched predictions published28
emo_ReliefIntense ReliefMatched predictions published64
emo_SadnessIntense SadnessMatched predictions published235
emo_Sexual_LustIntense Sexual LustMatched predictions published29
emo_ShameIntense ShameMatched predictions published28
emo_SournessIntense SournessMatched predictions published90
emo_TeasingIntense TeasingMatched predictions published110
emo_Thankfulness_GratitudeIntense Thankfulness GratitudeMatched predictions published87
emo_TriumphIntense TriumphMatched predictions published34
vn_S_ASMR_regStrong ASMR StyleMatched predictions published73
vn_S_AUTH_regStrong Authoritative StyleMatched predictions published209
vn_S_CART_regStrong Cartoonish StyleMatched predictions published33
vn_S_CASU_regStrong Casual StyleMatched predictions published154
vn_S_CONV_regStrong Conversational StyleMatched predictions published179
vn_S_DRAM_regStrong Dramatic StyleMatched predictions published445
vn_S_FORM_regStrong Formal StyleMatched predictions published174
vn_S_MONO_regStrong introspective monologue delivery, not monotone speechMatched predictions published223
vn_S_NARR_regStrong Narrator StyleMatched predictions published254
vn_S_NEWS_regStrong Newsreader StyleMatched predictions published28
vn_S_PLAY_regStrong Playful StyleMatched predictions published207
vn_S_RANT_regStrong Ranting/Angry StyleMatched predictions published31
vn_S_STRY_regStrong Storytelling StyleMatched predictions published127
vn_S_TECH_regStrong teacher-like, didactic deliveryMatched predictions published41
vn_S_WHIS_regStrong Whisper-Talk StyleMatched predictions published29
vn_VULN_reg_highemotionally vulnerable and exposedMatched predictions published318
vn_VULN_reg_lowemotionally detached and unexposedMatched predictions published237
vn_STNC_reg_highdominant and assertiveMatched predictions published222
vn_STNC_reg_lowsubmissive and deferentialMatched predictions published98
vn_AROU_reg_highhighly aroused and activatedMatched predictions published374
vn_AROU_reg_lowvery low arousal, calm or sleepyMatched predictions published83
vn_VOLT_reg_highEmotionally volatile delivery with rapid or abrupt changesMatched predictions published302
vn_VOLT_reg_lowEmotionally steady, stable delivery without abrupt changesMatched predictions published165
vn_VALN_reg_highstrongly positive vocal affectMatched predictions published185
vn_VALN_reg_lowstrongly negative vocal affectMatched predictions published325
vn_WARM_reg_highvery warm vocal deliveryMatched predictions published185
vn_WARM_reg_lowvery cold vocal deliveryOmitted by user request; no adapter0
vn_TENS_reg_highvery tense and strainedMatched predictions published352
vn_TENS_reg_lowvery relaxed and unstrainedMatched predictions published35
vn_TEMP_reg_highvery fast speaking tempoMatched predictions published101
vn_TEMP_reg_lowvery slow speaking tempoMatched predictions published29
casual_not_formalVery informal/casual, not formalMatched predictions published154
formal_not_casualVery formal, not casualMatched predictions published176

Data quality and reproducibility

Only individually audio-reviewed passing examples enter adapter training. The screened 45,112-clip pool is larger and not wholly verified. Empty/sparse buckets are not filled with weak audio and are not given fabricated adapters. Original source notices remain in the manually gated dataset; no new collection license is assigned.

Dataset, indices and taxonomy · Merged parent, adapters and code · Prompting manuals