Practical delivery recipes for all 40 emotions, 15 talking styles and 18 measurable vocal-burst classes, in English, German, French, Spanish and Dutch. Search results appear as workers finish; initial prompts are not presented as proven recommendations.
Model and inference code · Compact agent guide · Machine-readable recipes · Full experiment contract
Awaiting search: the experiment is queued or has no scored candidate yet. Provisional: a search candidate has scores but its independent check is unfinished. Held-out qualified: it passes the declared word-accuracy and holdout gates. Tested; no qualified recipe: the search finished but a recommendation would overstate its evidence.
All searches use the merged 1.7 model, without reference audio or extra adapters. A local language model mutates acting directions over three generations; spoken words stay fixed. The selected wording is checked at CFG 1–5 and on new texts and seeds. Higher CFG is not automatically better.
Emotion/style reward: ((2 × target + genuineness + burst blending) / 4) / (1 + WER). Each score uses its own frozen bounded normalization. Content Enjoyment is shown as a diagnostic and contributes zero to fitness. WER is a fraction, not a percentage. A take with WER > 15% or an unfinished output receives zero reward; failed seeds are never removed from averages.
Burst reward: confidence-weighted target-class time, measured by the independent locator and classifier in the actual between-sentence word gap, normalized to the requested burst duration, capped at one and penalized for extra bursts. It is also multiplied by 1 / (1 + WER). Each class is searched with and without a matching emotion. Shouted words alone are not a successful nonverbal scream. Detector recall varies sharply by class; a weak score can reflect measurement limits.
Three generations × nine strategies × six takes. The best wording then gets seven CFG settings, followed by six independent takes each for winner, runner-up and baseline. Timing conditions use both the round format seen in training and a legacy square format. Requested duration is conditioning, not a guarantee. Actual word-aligned timing errors are saved. Sealed texts are distinct within this search; historical EN/DE texts are not globally unseen.
Loading current results…
| Group / target | Language | Condition | Evidence | CFG / reward / WER |
|---|
Open any row for exact instructions, neutral CFG branch, candidate comparisons, all available audio and full scores.