Current steering-only ranking
EARS Steering v2 reuses the stored raw EARS predictions. Pure emotion/style conditions use only their named normalized head. Mixed scenes use the explicit head weights printed on each scenario and recipe. CE, genuineness, WER and speaker cosine are not weighted reward terms; WER/termination/identity and burst qualification remain separate gates. Original search fitness remains in a historical column. A new CPU ranking does not create an independent test of newly selected wording.
Immutable pilot, corrected follow-up
Pilot-v1 is a historical, two-generation developmental search. Its exact nested-duration prompts and event labels are preserved. The follow-up is a new experimental contract, not a silent repair of the old scores. Duration-only, label-only, and combined controls separate some format effects. Canonical repairs for non-event scenes omit redundant label controls.
Leading recipes are selected across both pilot generations. The final generation is not automatically best. New targeted interventions are a frozen ablation bank, not an additional unconstrained LLM evolution. Both models remain unchanged.
Historical search measures and unchanged gates
- EARS Medium predicts CE, genuineness, 40 emotion heads and VoiceNet attributes. These are model scores, not human MOS or calibrated probabilities.
- Targets use fixed EARS training mean/std CDF normalization; CE/10 and genuineness/6. Non-event fitness is 50% target + 25% CE + 25% genuine, multiplied by 1 − clipped WER.
- Legacy event fitness is 35% event + 20% coarse arc + 15% CE + 15% genuine + 15% blend, multiplied by 1 − clipped WER. New event-aware fitness retains that scalar separately and multiplies it by the fraction of requested families detected anywhere. A missing event cannot win on quality terms alone.
- ASR uses the frozen multilingual Parakeet model. Raw Unicode-aware WER is retained uncapped; clipping is only for the fitness factor.
- Speaker similarity is the pinned Orange 128-dimensional timbre embedding cosine. No universal nearest-neighbor 0.84 threshold is imported.
- Speech gates: per-take WER ≤ 15%, candidate mean WER ≤ 10%, valid/terminated fraction ≥ 90%. Reference mean cosine decline must be no worse than −0.05 with all 50 paired reference comparisons.
- New burst qualification additionally requires all requested families detected in at least 60% of takes. A detection anywhere does not prove correct onset. Literal interjection proxies are excluded from primary nonverbal recommendations.
Detector and timing limitations
The frozen vocal-burst locator and production x3 classifier have class-dependent recall and ambiguity between breaths, groans, screams and speech. Family and strict-label results differ. Nominal windows derive from requested timings, not actual aligned words. Chronological thirds provide only a coarse performance-arc diagnostic. A missed detection is not proof of no sound; a detected family is not proof of natural blending. Listen to examples and inspect reference/no-reference arms.
Whole generated waveforms, before MP3 encoding, supply search scores. Browser MP3s are lossy renderings of those same waveforms. The older release comparison uses its own published-MP3 metric contract and should not be numerically pooled without rescoring.
Selection, uncertainty and held-out evidence
Every 100-take cell contains ten scripts across two languages, five seeds and two reference arms. Complete distributions from baseline, leaders and runners-up are published as direct MP3s; complete packed evidence retains every candidate and failed outcome. No chosen best seed, restoration, denoising, voice conversion or time stretching is used.
The new transfer bank uses five unseen lexically neutral scripts per language. It tests the selected wording at its chosen CFG versus guidance 1. This is a limited held-out diagnostic, not independent human validation, a broad multilingual benchmark or evidence that every emotion/style works. Do not treat repeated language/reference variants as fully independent experiments.
Efficiency and reproducibility
Five seed replicas share one compatible inference batch with independent CUDA RNG streams. Pilot short-canary aggregate frame throughput improved approximately 4.5–5.4 times versus single generation; that is not an end-to-end guarantee. Follow-up jobs use 16 persistent generator/judge workers each, two 4-node jobs, with packed per-worker SQLite outputs. Historical controls are reused only after exact prompt, neutral prompt, seed and reference checks. No vast loose-file output tree is created on the cluster.
Audio and metadata are public under CC BY 4.0. Existing dependency/model licenses remain as stated in their repositories. All scenes are adult fiction. Use responsibly; do not represent synthetic performances as recordings of an unconsenting real person.