Fish Audio S2 Pro TTS

Fish Audio S2 Pro TTS.

Dual-AR architecture (Slow AR 4B + Fast AR 400M), 10 RVQ codebooks, ~21 Hz frame rate, 80+ languages.

• Model: fishaudio/s2-pro
• Output: 44.1 kHz WAV/PCM mono
• VRAM: ≥24GB for inference, A800/H200 recommended
• Technical Report: arXiv 2603.08823 | Architecture

Installation

See references/install.md. Quick summary:

bash

conda create -n fish-speech python=3.12 && conda activate fish-speech
pip install -e .[cu129]     # CUDA 12.9
# or: uv sync --python 3.12 --extra cu129
# minimal: pip install fish-speech

apt install portaudio19-dev libsox-dev ffmpeg  # System dependencies
hf download fishaudio/s2-pro --local-dir checkpoints/s2-pro

Server Deployment

vLLM-Omni (recommended, OpenAI compatible):

bash

pip install fish-speech
vllm serve fishaudio/s2-pro --omni --port 8091
# Endpoints: POST /v1/audio/speech, /v1/audio/speech/batch

SGLang-Omni (high-performance streaming):

bash

sgl-omni serve --model-path fishaudio/s2-pro --config examples/configs/s2pro_tts.yaml --port 8000
# RTF 0.195, TTFA ~100ms, throughput 3000+ t/s

Docker:

bash

docker compose --profile webui up    # Port 7860
COMPILE=1 docker compose --profile webui up  # ~10x speedup

Native API Server:

bash

python tools/api_server.py --llama-checkpoint-path checkpoints/s2-pro --decoder-checkpoint-path checkpoints/s2-pro/codec.pth --listen 0.0.0.0:8080

Raw CLI Inference (Three Steps)

bash

# 1. Extract VQ tokens
python fish_speech/models/dac/inference.py -i "ref.wav" --checkpoint-path "checkpoints/s2-pro/codec.pth"
# 2. Generate semantic tokens
python fish_speech/models/text2semantic/inference.py --text "Text" --prompt-text "Reference text" --prompt-tokens "fake.npy"
# 3. Decode to audio
python fish_speech/models/dac/inference.py -i "codes_0.npy"

API Calls

cURL

bash

# Basic TTS
curl -X POST http://localhost:8091/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello."}' --output out.wav

# Voice cloning (vLLM)
curl -X POST http://localhost:8091/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Cloned voice.", "ref_audio": "https://...", "ref_text": "Reference transcription"}' --output cloned.wav

# Streaming PCM
curl -N -X POST http://localhost:8091/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Streaming.", "stream": true, "response_format": "pcm"}' --no-buffer | play -t raw -r 44100 -e signed -b 16 -c 1 -

# Batch
curl -X POST http://localhost:8091/v1/audio/speech/batch \
  -H "Content-Type: application/json" \
  -d '{"items": [{"input": "Sentence 1"}, {"input": "Sentence 2"}], "voice": "default"}'

Python

python

import requests
resp = requests.post("http://localhost:8091/v1/audio/speech", json={
    "input": "Hello.", "voice": "default",
    "ref_audio": "https://...", "ref_text": "Reference text"
})
with open("out.wav", "wb") as f: f.write(resp.content)

# OpenAI SDK
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8091/v1", api_key="none")
client.audio.speech.create(model="fishaudio/s2-pro", voice="default", input="Hello.").stream_to_file("out.wav")

SGLang format: "references": [{"audio_path": "...", "text": "..."}]

Request Parameters

Parameter	Type	Default	Description
input	string	Required	Text to synthesize
voice	string	"default"	Voice
response_format	string	"wav"	wav/mp3/flac/pcm/aac/opus
speed	float	1.0	Speech speed (0.25-4.0)
stream	bool	false	Streaming (requires response_format="pcm")
ref_audio	string	null	Reference audio URL/base64/file://
ref_text	string	null	Reference audio transcription
max_new_tokens	int	2048	Max generation tokens
temperature	float	null	Sampling temperature
top_p	float	null	Nucleus sampling
top_k	int	null	Top-K
repetition_penalty	float	null	Repetition penalty
seed	int	null	Random seed

Emotion Tags

Embed [tag] anywhere in the text, supports 15000+ free-form tags:

text

[excited]Today is a great day![pause] [whisper in small voice]But there's a secret…
[professional broadcast tone]Welcome.

Common: [excited] [angry] [sad] [whisper] [shouting] [laughing] [pause] [emphasis] [echo] [inhale] [sigh] [singing]

Full reference: references/emotion-tags.md

Multi-Speaker

text

<|speaker:0|>Hello, welcome.
<|speaker:1|>Thank you, glad to be here.

LoRA Fine-tuning

⚠️ Not recommended for models after RL. Only fine-tune Slow AR:

bash

# Preparation: data/SPK1/*.mp3 + *.lab
python tools/vqgan/extract_vq.py data --config-name modded_dac_vq --checkpoint-path checkpoints/openaudio-s1-mini/codec.pth
python tools/llama/build_dataset.py --input data --output data/protos
python fish_speech/train.py --config-name text2semantic_finetune project=my_project [e-mail]_config=r_8_alpha_16
python tools/llama/merge_lora.py --lora-config r_8_alpha_16 --base-weight checkpoints/openaudio-s1-mini --lora-weight results/my_project/checkpoints/step_xxx.ckpt --output checkpoints/merged/

See references/finetune.md

Important Notes

1. Voice cloning: Reference audio 10-30 seconds, clear and noise-free, provide accurate transcription
2. Without reference audio, voice tends to sound mechanical
3. vLLM is easy to deploy; SGLang has better latency/throughput
4. SGLang: BF16 RoPE precision must match training; if early EOS occurs, switch to FA3
5. Fast AR torch.compile can achieve ~5x speedup
6. Docker image does not include model weights; mount checkpoints