Fish Audio S2 Pro is a state-of-the-art text-to-speech model with fine-grained inline control of prosody and emotion. Trained on 10M+ hours of audio data across 83 languages with 1500+ emotive tags, it combines reinforcement learning alignment with a Dual-Autoregressive architecture for speech that sounds natural, realistic, and emotionally rich.

Paper: Fish Audio S2 Technical Report (arXiv:2603.08823)

This ComfyUI wrapper provides native node-based integration with:

• Zero-shot voice cloning from 10-30 second reference audio
• Inline emotion/prosody control with [tag] syntax
• Multi-speaker conversation synthesis in a single pass
• Per-speaker audio isolation for multi-speaker lip sync workflows
• 83 language support with automatic detection

• ** Zero-Shot Voice Cloning** – Clone any voice from 10-30 seconds of reference audio
• ** 1500+ Emotive Tags** – Fine-grained control with [laugh], [whisper], [excited], [sad], etc.
• ** 83 Languages** – Full multilingual support without phoneme preprocessing
• ** Multi-Speaker TTS** – Generate conversations with multiple cloned voices in one pass
• ** Per-Speaker Audio Isolation** – Separate audio tracks for each speaker (lip sync workflows)
• ** Native ComfyUI Integration** – AUDIO noodle inputs, progress bars, interruption support
• ** Optimized Performance** – Support for bf16/fp16/fp32 dtypes, SDPA, FlashAttention, SageAttention
• ** Smart Auto-Download** – Model weights auto-downloaded from HuggingFace on first use
• ** Smart Caching** – Optional model caching with automatic unloading on config change

• GPU: NVIDIA GPU with 24GB+ VRAM for full model (RTX 3090/4090, A5000, etc.)
  • 16GB+ VRAM works with BNB NF4 4-bit on-the-fly quantization (~10-11 it/s)
  • CPU/MPS: ~1.5-2 seconds per token (experimental)
  • 18GB+ VRAM works with BNB INT8 on-the-fly quantization (~10-11 it/s)
  • 20GB+ VRAM works with the FP8 quantized model (s2-pro-fp8, ~15 it/s, requires RTX 4090/5090 or Ada/Blackwell GPU)
• CPU/MPS: ⚠️ EXPERIMENTAL
• Python: 3.10+
• CUDA: 11.8+ (for GPU inference)

⚠️ BNB On-the-Fly Quantization Requirements:

BNB INT8 and BNB NF4 options use the s2-pro (bf16) model and quantize on-the-fly via bitsandbytes.

Install bitsandbytes:

pip install bitsandbytes

Note: BNB options run at ~10-11 it/s vs ~15 it/s for FP8. They work on any NVIDIA GPU without special hardware requirements.

Models are auto-downloaded from HuggingFace on first use:

• fishaudio/s2-pro — full model
• drbaph/s2-pro-fp8 — FP8 quantized

Tested and working v0.4.0 with PyTorch 2.10+cu13.

As of v0.3.0, descript-audio-codec, descript-audiotools, and protobuf are never installed or modified by pip install -r requirements.txt. The two audio packages are auto-installed at first startup with --no-deps, leaving your environment's protobuf version untouched.

As of v0.3.6, all transitive runtime dependencies of dac/audiotools (flatten-dict, importlib-resources, julius, randomname, ffmpy, argbind) are also auto-installed, fixing fresh-install failures on clean portable environments.

cd ComfyUI/custom_nodes
git clone https://github.com/saganaki22/ComfyUI-FishAudioS2.git
cd ComfyUI-FishAudioS2
pip install -r requirements.txt

1. Download Model

   • Models are auto-downloaded from fishaudio/s2-pro on first use
   • Or manually download and place in ComfyUI/models/fishaudioS2/

2. Text-to-Speech

   • Add Fish S2 TTS node
   • Enter text with optional emotion tags: Hello! [excited] This is Fish Audio S2.
   • Select language (or use auto)
   • Run!

3. Voice Cloning

   • Add Fish S2 Voice Clone TTS node
   • Connect reference audio (10-30 seconds recommended)
   • Enter text to synthesize in cloned voice
   • Run!

4. Multi-Speaker

   • Add Fish S2 Multi-Speaker TTS node
   • Set number of speakers (2-10)
   • Connect reference audio for each speaker
   • Use [speaker_1]:, [speaker_2]: tokens in text
   • Run!

Text-to-speech synthesis with inline emotion/prosody control.

Inputs:

• model_path: S2-Pro checkpoint folder (place in ComfyUI/models/fishaudioS2/)
• text: Text to synthesize with optional [tag] emotion markers
• language: Language hint (auto, en, zh, ja, ko, etc.)
• device: Compute device (auto, cuda, cpu, mps)
• precision: Model precision (bfloat16, float16, float32)
• attention: Attention kernel (auto, sdpa, sage_attention, flash_attention)
• max_new_tokens: Maximum acoustic tokens (0 = auto)
• chunk_length: Chunk length (100-400) [Will be removed in future update]
• temperature: Sampling temperature (0.1-1.0)
• top_p: Top-p nucleus sampling (0.1-1.0)
• repetition_penalty: Repetition penalty (0.9-2.0)
• seed: Random seed
• keep_model_loaded: Cache model in VRAM between runs
• compile_model: Enable torch.compile (Linux only)

Outputs:

• audio: Generated speech (AUDIO)

Voice cloning from reference audio.

Inputs:

• All inputs from Fish S2 TTS, plus:
• reference_audio: Reference audio to clone (10-30 seconds recommended)
• reference_text (optional): Transcript of reference audio for improved accuracy

  Note: [Whisper (MTB)](https://github.com/mel Mass/mtb_nodes) does not work with newer transformers versions and will throw an error. Manually write the reference text until the MTB maintainer merges the fix PR.

Outputs:

• audio: Generated speech in cloned voice (AUDIO)

Multi-speaker conversation synthesis.

Inputs:

• All inputs from Fish S2 TTS, plus:
• num_speakers: Number of speakers (2-10)
• speaker_N_audio: Reference audio for speaker N
• speaker_N_ref_text: Optional transcript for speaker N

  Note: Whisper (MTB) does not work with newer transformers versions and will throw an error. Manually write the reference text until the MTB maintainer merges the fix PR.

Text Format:

[speaker_1]: Hello, I'm speaker one.
[speaker_2]: And I'm speaker two!

Outputs:

• audio: Generated multi-speaker conversation (AUDIO)

Multi-speaker conversation synthesis with separate audio tracks for each speaker. Perfect for multi-speaker lip sync workflows (e.g., Infinite Talk).

Inputs:

• Same as Fish S2 Multi-Speaker TTS

Text Format:

[speaker_1]: Hello, I'm speaker one.
[speaker_2]: And I'm speaker two!

Outputs:

• audio: Combined multi-speaker conversation (AUDIO)
• speaker_1_audio through speaker_10_audio: Isolated audio for each speaker (AUDIO)
  • Each speaker's track contains their speech when talking, silence otherwise
  • All tracks are the same length as the combined audio

Use Case: Connect individual speaker outputs to separate lip sync nodes for multi-character animation.

Credit: Per-speaker audio isolation idea suggested by @lazybuttalented. This is an independent implementation.

S2 Pro supports 1500+ unique emotive tags using [tag] syntax. These are free-form natural language descriptions, not predefined tags.

Common tags:

Free-form examples:

• [whisper in small voice]
• [super happy and excited]
• [speaking slowly and clearly]
• [sarcastic tone]

83 languages supported without phoneme preprocessing:

Tier 1 (Best Quality): Japanese (ja), English (en), Chinese (zh)

Tier 2: Korean (ko), Spanish (es), Portuguese (pt), Arabic (ar), Russian (ru), French (fr), German (de)

Full List: sv, it, tr, no, nl, cy, eu, ca, da, gl, ta, hu, fi, pl, et, hi, la, ur, th, vi, jw, bn, yo, sl, cs, sw, nn, he, ms, uk, id, kk, bg, lv, my, tl, sk, ne, fa, af, el, bo, hr, ro, sn, mi, yi, am, be, km, is, az, sd, br, sq, ps, mn, ht, ml, sr, sa, te, ka, bs, pa, lt, kn, si, hy, mr, as, gu, fo

ComfyUI/
├── models/
│   └── fishaudioS2/
│       └── s2-pro/                    # Full model (auto-downloaded)
│           ├── model.pt
│           └── config.json
└── custom_nodes/
    └── ComfyUI-FishAudioS2/
        ├── __init__.py
        ├── nodes/
        │   ├── tts_node.py
        │   ├── voice_clone_node.py
        │   ├── multi_speaker_node.py
        │   ├── loader.py
        │   └── model_cache.py
        ├── fish_speech_src/           # Bundled fish-speech source
        ├── requirements.txt
        └── README.md

pip install -U huggingface_hub
huggingface-cli download fishaudio/s2-pro --local-dir ComfyUI/models/fishaudioS2/s2-pro

huggingface-cli download drbaph/s2-pro-fp8 --local-dir ComfyUI/models/fishaudioS2/s2-pro-fp8

pip install descript-audio-codec --no-deps
pip install "descript-audiotools>=0.7.2" --no-deps

pip install -r requirements.txt
pip install flatten-dict importlib-resources julius randomname ffmpy argbind
pip install descript-audio-codec --no-deps
pip install "descript-audiotools>=0.7.2" --no-deps

• Model (Full): fishaudio/s2-pro
• Model (FP8 Quantized): drbaph/s2-pro-fp8
• Paper: huggingface.co/papers/2603.08823

• arXiv Paper: arxiv.org/abs/2603.08823
• Official Repository: fishaudio/fish-speech
• Documentation: speech.fish.audio

• Playground: fish.audio
• Discord: Fish Audio Discord
• Blog: Fish Audio S2 Release

If you use Fish Audio S2 in your research, please cite:

@misc{liao2026fishaudios2technical,
      title={Fish Audio S2 Technical Report}, 
      author={Shijia Liao and Yuxuan Wang and Songting Liu and Yifan Cheng and Ruoyi Zhang and Tianyu Li and Shidong Li and Yisheng Zheng and Xingwei Liu and Qingzheng Wang and Zhizhuo Zhou and Jiahua Liu and Xin Chen and Dawei Han},
      year={2026},
      eprint={2603.08823},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2603.08823},
}

This project uses the Fish Audio Research License. Research and non-commercial use is permitted. Commercial use requires a separate license from Fish Audio — contact business@fish.audio.

Model weights from fishaudio/s2-pro are subject to the same license.

Fish Audio S2 is intended for academic research, educational purposes, and legitimate applications. Please use responsibly and ethically. We do not hold any responsibility for any illegal usage. Please refer to your local laws about DMCA and related regulations.