Emotion × Voice · v2 — Delivery, Edge Cases & Styles

The setup

MOSS-VA-v2 (laion/moss-tts-local-transformer-4.55b-voice-acting-v2) is a 4.55B expressive text-to-speech model that can clone a voice from a few seconds of reference audio and follow a written stage direction. Per-emotion LoRA adapters nudge the voice toward a target feeling; a dial λ controls the dose. Each of 6 reference voices speaks 40 emotions in 4 delivery conditions — A intense & free, B moderate & free, C intense but contained, D moderate & contained.

What's new in v2

1. A hard speaker floor. Every winner must have ECAPA speaker similarity ≥ 0.50 to its reference — candidates below it are rejected; if none of 32 passes, the best available take is shown but flagged red (113 of 1032 cells). No more "different person" outliers. 2. Harder intensities. Intense conditions now merge the emotion LoRA at λ up to 1.9 (was 1.5) with more extreme stage directions for the high-arousal emotions (Fear, Anger, Pain, Distress, Triumph…). 3. Longer sentences. Every carrier sentence rewritten to ≥ 20 words (median 27) so there is room to act. 4. Edge cases. 8 vocal-burst situations per voice — fear/pain screams, groaning, shivering, laughter, crying, whimpering, tears — reproduced from our evolutionary edge-case search recipes, best-of-32 under the same floor. 5. Delivery styles. Very-quiet whispering, ASMR whispering, ranting and shouting via VoiceNet-dimension LoRA blends.

Selection

Per cell 32 candidates are generated; ranked by reward = z(genuineness)+z(vocal-burst-blend)+1.25·emotion-strength; the top-8 are checked against the speaker floor and the highest-reward take that still sounds like the speaker wins.