laion/humaneness-voice-small-acting-eval

Humaneness Voice Small Acting Evaluation

This release contains 18720 generated evaluation takes across 10 stages. See manifest.json for exact counts, hashes, and shard inventory.

Audio is stored in TAR shards. index.sqlite maps each take and audio role to a shard byte offset, length, checksum, and codec extension. Stage-level per-take metadata is available under exports/ in Parquet and JSONL form.

Scorer values are estimates and should not be interpreted as ground truth about emotion, authenticity, or audio quality. Read the accompanying protocol, scorer cards, and verified license information before reuse. The current challenge source is pinned to ad1886f7fec60b46bf81d46967dc8142fafc39e0. This card deliberately makes no license assertion for model-generated audio; a release owner must confirm the applicable terms before publication.

Postprocess metrics

Exact-class and relaxed-family burst metrics, per-sentence alignment, and duration-response results are attached to matched take records when available. Missing files are marked pending in postprocess_status; raw postprocess Parquet inputs are packed in artifacts/postprocess_metrics.tar when available. No sentence-timing value is inferred from clip-level diagnostics.

Frozen protocol and benchmark

The exact 50-scene challenge manifest, 11 prompt surfaces, 72 duration controls, reference-bank provenance, and the detailed English protocol are under protocol/. Their file hashes are recorded in manifest.json. The authored scene prompts are style-matched to training annotation surfaces, not newly generated by those Whisper/Gemma models.

Versioned target-aware acting reward

This per-take descriptive scalar is present only in a full audited postprocessed release. Version humaneness-voice-small-acting-reward-v3-canonical-burst-negative-filter uses R_base = (0.30 E_target + 0.20 V_target + 0.25 Q + 0.15 B + 0.10 D) / sum(applicable weights), then R_eval = R_base × max(0, 1 − WER) × G_empty. G_empty is zero for empty normalized ASR or a failed take, otherwise one; missing WER leaves the reward null. Missing applicable components also leave it null. Weights are renormalized only for genuinely nonapplicable components. E_target averages frozen-ECDF corrected historical intensity-band scores for requested heads and subtracts 0.25 times the mean excess of non-target emotion percentiles above 0.98. V_target averages linear target attainment, max(0, 1 − normalized absolute error/0.15), over explicitly requested dimensions. Q is the mean frozen corpus ECDF percentile of genuineness and, only for an explicit C01 burst prompt with canonical requested-type event F1 above zero, burst blend; otherwise Q uses genuineness alone. B is canonical requested-type plus onset event F1 from the burst postpass only for explicit C01 burst prompts. D averages max(0, 1 − sentence absolute duration error/requested sentence duration) only for timed prompts. The five-class canonical crosswalk is in protocol/burst_canonical_crosswalk_v2.json; unmatched detector event labels remain false positives; the detector's explicit no_burst negative/background class is excluded from events. Calibration path/SHA-256, scorer schema SHA-256, per-component values, applicability, weights, gate values, and missing-component reasons are recorded in the release/take data. The frozen corpus ECDF calibrates emotion, genuineness and blend; evaluation rows never fit calibration. This scalar is for descriptive sorting only and does not select checkpoints or replace the raw metric dashboard.