SIGN IN SIGN UP

fix(qwen3_tts): window ICL repetition penalty to stop pace acceleration

The ICL voice-cloning path forces repetition_penalty to a 1.5 floor and
penalized the set of all previously generated codec tokens. A slow
reference voice's recurring tokens stayed suppressed for the whole
utterance, progressively accelerating the speaking pace on long texts.

Apply the penalty over a bounded recent window (repetition_context_size,
default 64) in both _sample_token and _sample_token_batch, keeping the
penalty stationary while still catching short-range code degeneration.

Fixes #910

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A
Alazar Manakelew committed
d0e50f5b6f78197f479ee9494a30cf8195b9dc5b
Parent: 4da826f