[Feature] 新的TTS：Fish Audio S2

[zpengcom]

确认

• 我已确认这是一个新功能建议（而非 Bug 反馈），且之前无人提出

功能描述 / 痛点

模型介绍：https://github.com/fishaudio/fish-speech/blob/main/docs/README.zh.md

可用模型：
fishaudio/s2-pro — full model
baicai1145/s2-pro-w4a16 — GPTQ 4-bit quantized
drbaph/s2-pro-fp8 — FP8 quantized

演示demo：https://fish.audio/

参考资料或实现思路 (可选)

No response

[zhzLuke96]

好，这个模型看着不错

好像还支持多说话人？我得研究研究，多说话人的模型还没接入过

因为之前的多说话人模型都很大没试过，并且和整个架构不太匹配，这个估计可以试试

[zpengcom]

好，这个模型看着不错

好像还支持多说话人？我得研究研究，多说话人的模型还没接入过

因为之前的多说话人模型都很大没试过，并且和整个架构不太匹配，这个估计可以试试

支持多人多语言、 支持的语言超多，支持的标签也超多，实测效果也非常不错；
唯一问题非常吃资源。。看着模型很小，我4080也只能流畅跑bnb nf4

[JiwaniZakir]

Integrating Fish Audio S2 would require implementing a new TTS adapter under the existing engine abstraction (similar to how other backends like CosyVoice or ChatTTS are wrapped). The fish-speech architecture is notably different — it uses a dual-stage pipeline (AR language model → vocoder/decoder), so the adapter would need to handle the fish_speech inference flow including the codec tokenizer, rather than fitting a simpler end-to-end TTS pattern. The GPTQ and FP8 quantized variants (s2-pro-w4a16, s2-pro-fp8) would also need separate loading paths or at minimum verified compatibility with the existing quantization handling in the model loader. Worth checking whether the existing TTSEngine base class interface accommodates streaming token generation, since fish-speech's AR stage naturally supports it and bypassing that would be a performance regression.