Most open text-to-speech projects treat emotion as an afterthought — a neutral voice with maybe a speed slider. Fish Speech inverts that priority: prosody and emotion are steerable at the sub-word level through inline natural-language tags, so a single sentence can slide from [whisper] to [excited] without re-recording. Under the hood sits Fish Audio S2 Pro, a Dual-Autoregressive model trained on 10M+ hours of audio across 80+ languages with reinforcement-learning alignment.
What Sets It Apart
- Sub-word emotional control — tags like
[angry]or[whisper]apply mid-sentence, giving direction-level nuance instead of one flat delivery per clip. - Native multi-speaker, multi-turn dialogue — generate whole conversations or character audio in a single pass rather than stitching separate renders.
- Serving-ready by design — CLI, WebUI, and server inference plus SGLang-Omni and vLLM-Omni recipes mean it drops into production stacks, not just notebooks.
- Breadth over 80 languages — the same model handles cross-lingual cloning, so you are not swapping engines per locale.
Who It's For
Great fit if you need expressive, controllable multilingual TTS or voice cloning and can operate within a research-oriented license. Look elsewhere if you need clean commercial redistribution rights out of the box — the Fish Audio Research License restricts how weights and outputs may be used, so read it before shipping anything customer-facing.
