fishaudio/fish-speech

Implements a Dual-Autoregressive architecture combining a 4B-parameter slow decoder with a 400M fast decoder for semantic and acoustic codebook generation, trained on 10M+ hours across 80+ languages. Supports sub-word prosody and emotion control via inline natural language tags (e.g., `[whisper]`, `[excited]`), enabling multi-speaker conversations with reinforcement learning alignment for instruction adherence and naturalness.

26,613 stars. Actively maintained with 26 commits in the last 30 days.