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.
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.
Best tested agent route: Use manual-only configuration; no clear qualifying incremental adapter gain. Bucket scale 0.0, CFG 4.0.
| Held-out condition | n | Target z↑ | Raw head | CE↑ | Genuine↑ | Voice↑ | WER↓ | Valid | QA |
|---|---|---|---|---|---|---|---|---|---|
| standard | 50 | 1.190 | 1.143 | 6.214 | 4.343 | 0.577 | 0.024 | 0.980 | pass |
| manual | 50 | 2.083 | 1.635 | 6.274 | 4.123 | 0.543 | 0.012 | 0.980 | pass |
| adapter | 50 | 2.257 | 1.731 | 6.240 | 4.039 | 0.545 | 0.017 | 1.000 | pass |
| matched_no_adapter | 50 | 2.289 | 1.749 | 6.223 | 4.057 | 0.557 | 0.008 | 1.000 | pass |
Incremental adapter gain versus same-CFG control: -0.032 z;95% cluster interval [-0.1382031033735609, 0.09523443945676088]. Versus best manual-only: 0.174 z. Manual-only versus standard: 0.893 z.
| Scale | CFG | Target z | CE | Genuine | WER | Provisional QA |
|---|---|---|---|---|---|---|
| 0.0 | 1.0 | 1.352 | 6.211 | 4.257 | 0.000 | True |
| 0.5 | 1.0 | 1.252 | 6.213 | 4.341 | 0.021 | True |
| 0.75 | 1.0 | 1.233 | 6.224 | 4.280 | 0.000 | True |
| 1.0 | 1.0 | 1.210 | 6.241 | 4.335 | 0.000 | True |
| 0.0 | 1.25 | 1.300 | 6.212 | 4.278 | 0.000 | True |
| 0.5 | 1.25 | 1.527 | 6.195 | 4.235 | 0.000 | True |
| 0.75 | 1.25 | 1.390 | 6.251 | 4.301 | 0.003 | True |
| 1.0 | 1.25 | 1.315 | 6.245 | 4.317 | 0.000 | True |
| 0.0 | 1.5 | 1.317 | 6.183 | 4.302 | 0.003 | True |
| 0.5 | 1.5 | 1.591 | 6.227 | 4.237 | 0.000 | True |
| 0.75 | 1.5 | 1.217 | 6.224 | 4.287 | 0.000 | True |
| 1.0 | 1.5 | 1.356 | 6.223 | 4.301 | 0.000 | True |
| 0.0 | 2.0 | 1.375 | 6.205 | 4.281 | 0.000 | True |
| 0.5 | 2.0 | 1.511 | 6.212 | 4.252 | 0.000 | True |
| 0.75 | 2.0 | 1.650 | 6.181 | 4.258 | 0.000 | True |
| 1.0 | 2.0 | 1.222 | 6.222 | 4.327 | 0.000 | True |
| 0.0 | 2.5 | 1.579 | 6.201 | 4.213 | 0.005 | True |
| 0.5 | 2.5 | 1.725 | 6.219 | 4.185 | 0.000 | True |
| 0.75 | 2.5 | 1.485 | 6.224 | 4.269 | 0.000 | True |
| 1.0 | 2.5 | 1.663 | 6.200 | 4.223 | 0.003 | True |
| 0.0 | 3.0 | 1.770 | 6.199 | 4.191 | 0.003 | True |
| 0.5 | 3.0 | 1.790 | 6.210 | 4.184 | 0.003 | True |
| 0.75 | 3.0 | 1.577 | 6.182 | 4.189 | 0.034 | True |
| 1.0 | 3.0 | 1.799 | 6.181 | 4.128 | 0.003 | True |
| 0.0 | 4.0 | 1.784 | 6.218 | 4.178 | 0.000 | True |
| 0.5 | 4.0 | 1.540 | 6.199 | 4.162 | 0.006 | True |
| 0.75 | 4.0 | 1.829 | 6.176 | 4.104 | 0.000 | True |
| 1.0 | 4.0 | 1.763 | 6.201 | 4.162 | 0.000 | True |
| 0.0 | 5.0 | 1.734 | 6.221 | 4.177 | 0.021 | True |
| 0.5 | 5.0 | 1.858 | 6.190 | 4.106 | 0.000 | True |
| 0.75 | 5.0 | 1.815 | 6.189 | 4.105 | 0.000 | True |
| 1.0 | 5.0 | 1.878 | 6.182 | 4.114 | 0.000 | True |
direct_manual_recipe
{
"id": "manual-59e65fa10d8f",
"general": "Deliver with a highly polished, almost academic neutrality, prioritizing impeccable articulation.",
"inline": [
"Start with a crisp, clear articulation that establishes immediate factual distance.",
"Ensure the pace remains consistently regulated, avoiding any natural speech ebb and flow.",
"Preserve this highly controlled, precise delivery right up to the very last sound."
],
"cfg": 3,
"recipe": "P06",
"strategy": "local-Gemma-mutation",
"timing_syntax": "canonical",
"canonical_labels": true,
"neutral_event": "keep",
"parent": "cfg-3",
"origin": "exact_frozen_manual_recipe"
}Exact machine-readable comparison · Agent tool recommendations
Strong Newsreader Style. Base: 1.5 Alpha + merged real-speech rank4; new adapter: rank32. Every row is a matched seed/text/reference test. Scale zero disables the new adapter; all other scales are applied absolutely, never compounded.
Adapter training data: {"es": 6, "nl": 5, "fr": 8, "en": 7, "de": 9}. All available accepted languages were pooled, globally shuffled, for three epochs. Peak LR 2e-05 at the end of the first-epoch warmup, then cosine decay. All original optimizer/RNG checkpoints remain saved.
Taxonomy: S_MONO is monologue, not monotone; VOLT is volatility, not volume. No claim is made that a classifier score alone proves the audible performance.
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 20 | 0.815 ± 0.170 | 6.208 ± 0.169 | 4.369 ± 0.222 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 20 | 0.788 ± 0.153 | 6.179 ± 0.175 | 4.376 ± 0.274 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 20 | 0.815 ± 0.129 | 6.158 ± 0.154 | 4.294 ± 0.145 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 20 | 0.794 ± 0.161 | 6.190 ± 0.136 | 4.359 ± 0.262 | 0.00% ± 0.00 |
Fixed speaker reference:
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 4 | 0.625 ± 0.169 | 6.350 ± 0.040 | 4.366 ± 0.288 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 4 | 0.590 ± 0.121 | 6.307 ± 0.142 | 4.244 ± 0.326 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 4 | 0.659 ± 0.077 | 6.281 ± 0.123 | 4.264 ± 0.162 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 4 | 0.563 ± 0.150 | 6.271 ± 0.102 | 4.252 ± 0.426 | 0.00% ± 0.00 |
| Prompt / seed | Merged base · no new LoRA | Rank32 × 0.5 | Rank32 × 0.75 | Rank32 × 1.0 |
|---|---|---|---|---|
| Seed 1701 The door is still open. I put your coat on the chair. Tell me when you are ready. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] The door is still open. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] I put your coat on the chair. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Tell me when you are ready. | Target 0.785 · CE 6.30 · Genuine 4.23 · WER 0.0% | Target 0.533 · CE 6.36 · Genuine 4.24 · WER 0.0% | Target 0.634 · CE 6.18 · Genuine 4.22 · WER 0.0% | Target 0.533 · CE 6.31 · Genuine 4.21 · WER 0.0% |
| Seed 1702 The door is still open. I put your coat on the chair. Tell me when you are ready. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] The door is still open. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] I put your coat on the chair. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Tell me when you are ready. | Target 0.752 · CE 6.34 · Genuine 4.03 · WER 0.0% | Target 0.764 · CE 6.11 · Genuine 3.82 · WER 0.0% | Target 0.739 · CE 6.25 · Genuine 4.06 · WER 0.0% | Target 0.783 · CE 6.17 · Genuine 3.67 · WER 0.0% |
| Seed 1701 I found the little box. It was beside the old window. We can look inside together. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] I found the little box. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] It was beside the old window. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] We can look inside together. | Target 0.441 · CE 6.36 · Genuine 4.66 · WER 0.0% | Target 0.491 · CE 6.44 · Genuine 4.61 · WER 0.0% | Target 0.698 · CE 6.46 · Genuine 4.41 · WER 0.0% | Target 0.481 · CE 6.21 · Genuine 4.62 · WER 0.0% |
| Seed 1702 I found the little box. It was beside the old window. We can look inside together. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] I found the little box. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] It was beside the old window. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] We can look inside together. | Target 0.523 · CE 6.40 · Genuine 4.55 · WER 0.0% | Target 0.573 · CE 6.31 · Genuine 4.31 · WER 0.0% | Target 0.563 · CE 6.23 · Genuine 4.37 · WER 0.0% | Target 0.455 · CE 6.40 · Genuine 4.51 · WER 0.0% |
Fixed speaker reference:
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 4 | 0.856 ± 0.227 | 6.326 ± 0.187 | 4.296 ± 0.238 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 4 | 0.877 ± 0.168 | 6.316 ± 0.067 | 4.443 ± 0.289 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 4 | 0.935 ± 0.037 | 6.283 ± 0.161 | 4.350 ± 0.123 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 4 | 0.944 ± 0.034 | 6.328 ± 0.113 | 4.354 ± 0.082 | 0.00% ± 0.00 |
| Prompt / seed | Merged base · no new LoRA | Rank32 × 0.5 | Rank32 × 0.75 | Rank32 × 1.0 |
|---|---|---|---|---|
| Seed 1701 Die Tür ist noch offen. Ich habe deinen Mantel auf den Stuhl gelegt. Sag mir, wenn du bereit bist. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Die Tür ist noch offen. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Ich habe deinen Mantel auf den Stuhl gelegt. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Sag mir, wenn du bereit bist. | Target 0.931 · CE 6.09 · Genuine 4.19 · WER 0.0% | Target 0.945 · CE 6.24 · Genuine 4.47 · WER 0.0% | Target 0.931 · CE 6.09 · Genuine 4.28 · WER 0.0% | Target 0.951 · CE 6.30 · Genuine 4.36 · WER 0.0% |
| Seed 1702 Die Tür ist noch offen. Ich habe deinen Mantel auf den Stuhl gelegt. Sag mir, wenn du bereit bist. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Die Tür ist noch offen. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Ich habe deinen Mantel auf den Stuhl gelegt. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Sag mir, wenn du bereit bist. | Target 0.518 · CE 6.28 · Genuine 4.65 · WER 0.0% | Target 0.627 · CE 6.28 · Genuine 4.83 · WER 0.0% | Target 0.884 · CE 6.21 · Genuine 4.53 · WER 0.0% | Target 0.904 · CE 6.18 · Genuine 4.34 · WER 0.0% |
| Seed 1701 Ich habe die kleine Schachtel gefunden. Sie stand neben dem alten Fenster. Wir können gemeinsam hineinschauen. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Ich habe die kleine Schachtel gefunden. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Sie stand neben dem alten Fenster. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Wir können gemeinsam hineinschauen. | Target 0.987 · CE 6.51 · Genuine 4.22 · WER 0.0% | Target 0.987 · CE 6.36 · Genuine 4.16 · WER 0.0% | Target 0.970 · CE 6.44 · Genuine 4.32 · WER 0.0% | Target 0.987 · CE 6.44 · Genuine 4.26 · WER 0.0% |
| Seed 1702 Ich habe die kleine Schachtel gefunden. Sie stand neben dem alten Fenster. Wir können gemeinsam hineinschauen. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Ich habe die kleine Schachtel gefunden. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Sie stand neben dem alten Fenster. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Wir können gemeinsam hineinschauen. | Target 0.987 · CE 6.44 · Genuine 4.12 · WER 0.0% | Target 0.948 · CE 6.38 · Genuine 4.31 · WER 0.0% | Target 0.954 · CE 6.40 · Genuine 4.27 · WER 0.0% | Target 0.935 · CE 6.40 · Genuine 4.46 · WER 0.0% |
Fixed speaker reference:
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 4 | 0.792 ± 0.156 | 6.231 ± 0.066 | 4.474 ± 0.221 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 4 | 0.794 ± 0.069 | 6.144 ± 0.106 | 4.476 ± 0.130 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 4 | 0.827 ± 0.061 | 6.176 ± 0.084 | 4.319 ± 0.070 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 4 | 0.817 ± 0.126 | 6.138 ± 0.101 | 4.575 ± 0.148 | 0.00% ± 0.00 |
| Prompt / seed | Merged base · no new LoRA | Rank32 × 0.5 | Rank32 × 0.75 | Rank32 × 1.0 |
|---|---|---|---|---|
| Seed 1701 La porte est encore ouverte. J’ai posé ton manteau sur la chaise. Dis-moi quand tu es prêt. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] La porte est encore ouverte. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] J’ai posé ton manteau sur la chaise. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Dis-moi quand tu es prêt. | Target 0.584 · CE 6.32 · Genuine 4.70 · WER 0.0% | Target 0.881 · CE 5.99 · Genuine 4.36 · WER 0.0% | Target 0.740 · CE 6.26 · Genuine 4.41 · WER 0.0% | Target 0.691 · CE 6.05 · Genuine 4.63 · WER 0.0% |
| Seed 1702 La porte est encore ouverte. J’ai posé ton manteau sur la chaise. Dis-moi quand tu es prêt. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] La porte est encore ouverte. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] J’ai posé ton manteau sur la chaise. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Dis-moi quand tu es prêt. | Target 0.927 · CE 6.25 · Genuine 4.29 · WER 0.0% | Target 0.809 · CE 6.17 · Genuine 4.39 · WER 0.0% | Target 0.873 · CE 6.17 · Genuine 4.26 · WER 0.0% | Target 0.930 · CE 6.07 · Genuine 4.40 · WER 0.0% |
| Seed 1701 J’ai trouvé la petite boîte. Elle était près de la vieille fenêtre. Nous pouvons regarder ensemble ce qu’elle contient. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] J’ai trouvé la petite boîte. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Elle était près de la vieille fenêtre. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Nous pouvons regarder ensemble ce qu’elle contient. | Target 0.760 · CE 6.17 · Genuine 4.63 · WER 0.0% | Target 0.718 · CE 6.24 · Genuine 4.65 · WER 0.0% | Target 0.829 · CE 6.21 · Genuine 4.34 · WER 0.0% | Target 0.921 · CE 6.17 · Genuine 4.52 · WER 0.0% |
| Seed 1702 J’ai trouvé la petite boîte. Elle était près de la vieille fenêtre. Nous pouvons regarder ensemble ce qu’elle contient. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] J’ai trouvé la petite boîte. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Elle était près de la vieille fenêtre. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Nous pouvons regarder ensemble ce qu’elle contient. | Target 0.897 · CE 6.20 · Genuine 4.28 · WER 0.0% | Target 0.769 · CE 6.17 · Genuine 4.51 · WER 0.0% | Target 0.865 · CE 6.06 · Genuine 4.27 · WER 0.0% | Target 0.726 · CE 6.27 · Genuine 4.75 · WER 0.0% |
Fixed speaker reference:
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 4 | 0.906 ± 0.056 | 6.015 ± 0.162 | 4.365 ± 0.191 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 4 | 0.778 ± 0.064 | 6.168 ± 0.047 | 4.399 ± 0.041 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 4 | 0.764 ± 0.130 | 6.031 ± 0.094 | 4.291 ± 0.117 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 4 | 0.766 ± 0.083 | 6.094 ± 0.142 | 4.283 ± 0.164 | 0.00% ± 0.00 |
| Prompt / seed | Merged base · no new LoRA | Rank32 × 0.5 | Rank32 × 0.75 | Rank32 × 1.0 |
|---|---|---|---|---|
| Seed 1701 De deur staat nog open. Ik heb je jas op de stoel gelegd. Zeg maar wanneer je klaar bent. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] De deur staat nog open. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Ik heb je jas op de stoel gelegd. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Zeg maar wanneer je klaar bent. | Target 0.878 · CE 5.97 · Genuine 4.20 · WER 0.0% | Target 0.765 · CE 6.13 · Genuine 4.38 · WER 0.0% | Target 0.683 · CE 5.99 · Genuine 4.34 · WER 0.0% | Target 0.649 · CE 5.94 · Genuine 4.04 · WER 0.0% |
| Seed 1702 De deur staat nog open. Ik heb je jas op de stoel gelegd. Zeg maar wanneer je klaar bent. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] De deur staat nog open. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Ik heb je jas op de stoel gelegd. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Zeg maar wanneer je klaar bent. | Target 0.905 · CE 5.82 · Genuine 4.61 · WER 0.0% | Target 0.693 · CE 6.16 · Genuine 4.39 · WER 0.0% | Target 0.721 · CE 5.95 · Genuine 4.39 · WER 0.0% | Target 0.766 · CE 6.11 · Genuine 4.36 · WER 0.0% |
| Seed 1701 Ik heb het kleine doosje gevonden. Het stond naast het oude raam. We kunnen er samen in kijken. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Ik heb het kleine doosje gevonden. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Het stond naast het oude raam. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] We kunnen er samen in kijken. | Target 0.856 · CE 6.20 · Genuine 4.43 · WER 0.0% | Target 0.836 · CE 6.14 · Genuine 4.46 · WER 0.0% | Target 0.693 · CE 6.17 · Genuine 4.31 · WER 0.0% | Target 0.832 · CE 6.28 · Genuine 4.37 · WER 0.0% |
| Seed 1702 Ik heb het kleine doosje gevonden. Het stond naast het oude raam. We kunnen er samen in kijken. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Ik heb het kleine doosje gevonden. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Het stond naast het oude raam. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] We kunnen er samen in kijken. | Target 0.983 · CE 6.08 · Genuine 4.22 · WER 0.0% | Target 0.817 · CE 6.24 · Genuine 4.36 · WER 0.0% | Target 0.957 · CE 6.01 · Genuine 4.12 · WER 0.0% | Target 0.818 · CE 6.05 · Genuine 4.37 · WER 0.0% |
Fixed speaker reference:
| Condition | n | Target score ↑ | Content Enjoyment ↑ | Genuineness ↑ | WER ↓ |
|---|---|---|---|---|---|
| Merged base, no bucket adapter | 4 | 0.894 ± 0.063 | 6.116 ± 0.073 | 4.344 ± 0.255 | 0.00% ± 0.00 |
| Rank32 × 0.5 | 4 | 0.901 ± 0.129 | 5.963 ± 0.212 | 4.319 ± 0.468 | 0.00% ± 0.00 |
| Rank32 × 0.75 | 4 | 0.890 ± 0.120 | 6.019 ± 0.072 | 4.247 ± 0.252 | 0.00% ± 0.00 |
| Rank32 × 1.0 | 4 | 0.881 ± 0.064 | 6.120 ± 0.081 | 4.332 ± 0.331 | 0.00% ± 0.00 |
| Prompt / seed | Merged base · no new LoRA | Rank32 × 0.5 | Rank32 × 0.75 | Rank32 × 1.0 |
|---|---|---|---|---|
| Seed 1701 La puerta sigue abierta. He dejado tu abrigo sobre la silla. Dime cuándo estás listo. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] La puerta sigue abierta. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] He dejado tu abrigo sobre la silla. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Dime cuándo estás listo. | Target 0.905 · CE 6.14 · Genuine 4.08 · WER 0.0% | Target 0.993 · CE 5.79 · Genuine 3.79 · WER 0.0% | Target 0.950 · CE 5.99 · Genuine 3.94 · WER 0.0% | Target 0.964 · CE 6.13 · Genuine 3.96 · WER 0.0% |
| Seed 1702 La puerta sigue abierta. He dejado tu abrigo sobre la silla. Dime cuándo estás listo. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] La puerta sigue abierta. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] He dejado tu abrigo sobre la silla. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Dime cuándo estás listo. | Target 0.908 · CE 6.04 · Genuine 4.21 · WER 0.0% | Target 0.909 · CE 5.77 · Genuine 4.12 · WER 0.0% | Target 0.952 · CE 5.99 · Genuine 4.15 · WER 0.0% | Target 0.860 · CE 6.02 · Genuine 4.16 · WER 0.0% |
| Seed 1701 He encontrado la cajita. Estaba junto a la ventana vieja. Podemos mirar juntos lo que hay dentro. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] He encontrado la cajita. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Estaba junto a la ventana vieja. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Podemos mirar juntos lo que hay dentro. | Target 0.957 · CE 6.21 · Genuine 4.43 · WER 0.0% | Target 0.985 · CE 6.12 · Genuine 4.49 · WER 0.0% | Target 0.950 · CE 6.13 · Genuine 4.39 · WER 0.0% | Target 0.889 · CE 6.22 · Genuine 4.57 · WER 0.0% |
| Seed 1702 He encontrado la cajita. Estaba junto a la ventana vieja. Podemos mirar juntos lo que hay dentro. Exact GENERAL / SCRIPT promptGENERAL: One adult speaker in a clean recording. Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word. SCRIPT: (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] He encontrado la cajita. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.8 seconds duration] Estaba junto a la ventana vieja. (Strong Newsreader Style. Make the requested delivery clearly audible, not merely implied by the words. Preserve every word.) [3.2 seconds duration] Podemos mirar juntos lo que hay dentro. | Target 0.806 · CE 6.08 · Genuine 4.66 · WER 0.0% | Target 0.716 · CE 6.17 · Genuine 4.87 · WER 0.0% | Target 0.710 · CE 5.97 · Genuine 4.51 · WER 0.0% | Target 0.811 · CE 6.11 · Genuine 4.65 · WER 0.0% |