laion/humaneness-voice-small-dpo-acting-eval

Humaneness Voice Small DPO: matched acting evaluation audio

This is a matched, listenable five-way comparison: the original S3 base plus four S3 DPO LoRA adapters (original/rebalanced task mix × ranks 64/128). Each LoRA was evaluated on the exact frozen Humaneness Voice Small acting benchmark already used for S3: 50 scenes, 11 prompt conditions, both English and German, with the same scripts, prompts, reference assignments and generation seeds. There are 1,872 matched takes per model. The four LoRAs contribute 7,488 directly addressable 128-kbps mono MP3 files here; the S3 MP3s remain in the existing benchmark repository and are linked by the side-by-side listening Space.

The comparison key is (challenge_id, surface, mode, seed). The MP3 path for an adapter is mp3/{model}/{challenge_id}__{surface}__{mode}__seed{seed}.mp3, for example mp3/original_rank064/HVSAC-001__C01__instruction__seed17.mp3. Every MP3 is playable through the Hub's resolve/main/... URL with browser range requests. The 16 scores/{model}/rankNNN.parquet files retain one row per take and the complete Parakeet ASR/WER, genuineness, vocal-burst blending, Empathic Insight Voice Plus 40-emotion vector, VoiceNet vector, clip-duration diagnostics, prompt and provenance. proofs/ contains immutable per-rank PASS records and file hashes. summary.json reports matched-row statistics and missingness.

How to read the results

The Space shows only common-key matched comparisons, not unrelated random samples. Lower WER is better. Higher genuineness and blend predictor values usually indicate stronger model-predicted naturalness/blending, but predictors are imperfect and should not replace listening. The target-emotion activation is the mean output of the requested emotion heads for that scene; it is a predictor activation, not a calibrated emotion-accuracy percentage. Vocal-burst F1 is summarized only on scenes that request bursts. Clip-level duration error is a diagnostic only. Forced sentence alignment is pending, so this release does not claim that per-sentence timing compliance has been validated. Source PASS files can carry publication_ready: false for that missing alignment postpass even when generation and first-pass scoring succeeded. Compare prompts and audio yourself before selecting an adapter.

All four adapter checkpoints are derived from the same S3 base. The original DPO pair mix contains 715,402 pairs; the rebalanced six-task mix has 447,179. The rebalanced run has no independent pair-level pitch or quality holdout, so pair accuracy cannot prove those two additions improve audio. See the model card for full training caveats and reproducible code. The underlying acting-challenge prompts are from laion/acting-challenge-dataset, licensed Apache 2.0; generated audio and this derived evaluation release are CC BY 4.0, consistent with the S3 base model. Do not use this benchmark as a test of speaker consent or identity fidelity.