LAION · matched listening study
Humaneness Voice Small: S3 vs four DPO LoRAs
Five versions of the same acting challenge, prompt, reference assignment and seed—one original S3 base and four preference-tuned LoRA adapters. Browse 50 scenes, 11 prompt surfaces and 1,872 matched keys. Audio streams directly from Hugging Face MP3 URLs; only the currently visible rows create audio players.
Model and adapters · MP3s, raw score Parquets and provenance · Original S3 benchmark
WER, genuineness, blending, burst and emotion values are model-based measurements, not human ratings. Vocal-burst F1 applies only to burst prompts. Sentence-level forced alignment is still pending: clip-duration error below is diagnostic and is not proof of sentence timing accuracy.
Matched-set statistics
Loading scores…
| Model | WER ↓ | Genuineness ↑ | Vocal-burst blend ↑ | Requested-burst F1 ↑ | Prompted emotion-head activation ↑ | Clip duration error (s) ↓ |
|---|
Each cell reports mean ± standard deviation and its available n; a missing value is excluded and counted. Emotion activation averages the requested emotion predictor heads, not calibrated accuracy. Sort/listen below to judge whether a metric agrees with the sound.