moonbox-dev/s2-pro-MLX

Fish Audio S2 Pro MLX

MLX builds of Fish Audio S2 Pro, made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's files together. Only the parts listed below are converted: the rest of the pipeline comes from the original repository unless the section below says otherwise.

Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below for what was measured).

Precision	File	Size	sha256	Checked against the original
bf16	bf16/chat_template.jinja	4.1 kB	87a2728cb8dc	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/codec.safetensors	2.78 GB	fde06aa1f23b	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/config.json	2.1 kB	870046f617ca	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/model-00001-of-00002.safetensors	5.35 GB	13c1b614ac67	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/model-00002-of-00002.safetensors	3.78 GB	a75a5d2e91dd	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/model.safetensors.index.json	30.4 kB	65dc730b4e46	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/special_tokens_map.json	0.1 MB	c2ff18fde6e4	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/tokenizer.json	12.2 MB	f24e08099d45	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
bf16	bf16/tokenizer_config.json	0.9 MB	b8d149343ae4	WER 0.0 % / CER 0.0 %, speaker sim 0.956 / 0.952, DNSMOS 3.24 / 2.92 (original 0.0 % / 0.0 %, 0.952 / 0.949, 3.21 / 2.76)
fp16	fp16/chat_template.jinja	4.1 kB	87a2728cb8dc	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/codec.safetensors	2.78 GB	fde06aa1f23b	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/config.json	2.1 kB	870046f617ca	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/model-00001-of-00002.safetensors	5.35 GB	d70c8e348abd	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/model-00002-of-00002.safetensors	3.78 GB	94b99c94994c	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/model.safetensors.index.json	30.4 kB	65dc730b4e46	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/special_tokens_map.json	0.1 MB	c2ff18fde6e4	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/tokenizer.json	12.2 MB	f24e08099d45	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
fp16	fp16/tokenizer_config.json	0.9 MB	b8d149343ae4	WER 0.0 % / CER 0.0 %, speaker sim 0.964 / 0.951, DNSMOS 3.21 / 2.86
8bit	8bit/chat_template.jinja	4.1 kB	87a2728cb8dc	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/codec.safetensors	2.78 GB	fde06aa1f23b	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/config.json	2.3 kB	363895eff17c	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/model-00001-of-00002.safetensors	5.35 GB	7d96671b6003	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/model-00002-of-00002.safetensors	372.3 MB	876e059805ce	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/model.safetensors.index.json	61.0 kB	7815620c53eb	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/special_tokens_map.json	0.1 MB	c2ff18fde6e4	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/tokenizer.json	12.2 MB	f24e08099d45	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
8bit	8bit/tokenizer_config.json	0.9 MB	b8d149343ae4	WER 0.0 % / CER 0.0 %, speaker sim 0.953 / 0.953, DNSMOS 3.22 / 2.83
6bit	6bit/chat_template.jinja	4.1 kB	87a2728cb8dc	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/codec.safetensors	2.78 GB	fde06aa1f23b	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/config.json	2.3 kB	e26146aa01b3	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/model.safetensors	4.81 GB	1c8b473cc7a9	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/model.safetensors.index.json	50.2 kB	dc9f4a81f88b	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/special_tokens_map.json	0.1 MB	c2ff18fde6e4	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/tokenizer.json	12.2 MB	f24e08099d45	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
6bit	6bit/tokenizer_config.json	0.9 MB	b8d149343ae4	WER 0.0 % / CER 0.0 %, speaker sim 0.966 / 0.945, DNSMOS 3.24 / 2.80
4bit	4bit/chat_template.jinja	4.1 kB	87a2728cb8dc	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/codec.safetensors	2.78 GB	fde06aa1f23b	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/config.json	2.3 kB	04a6afe6a550	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/model.safetensors	3.90 GB	f57eb1c13b76	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/model.safetensors.index.json	50.2 kB	5c1dd949a15f	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/special_tokens_map.json	0.1 MB	c2ff18fde6e4	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/tokenizer.json	12.2 MB	f24e08099d45	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01
4bit	4bit/tokenizer_config.json	0.9 MB	b8d149343ae4	WER 0.0 % / CER 0.0 %, speaker sim 0.963 / 0.944, DNSMOS 3.22 / 3.01

How to run

Fish Audio S2 Pro (4.4B dual-autoregressive TTS: voice cloning, 80+ languages, inline [tag] control of emotion and prosody) for mlx-audio. Built with Fish Audio. Each folder is a complete mlx-audio model: model-*.safetensors (slow and fast AR transformers), codec.safetensors (the DAC codec, float32, re-saved unchanged from the original codec.pth), the tokenizer and chat template. Tested with mlx-audio==0.6.5:

from mlx_audio.tts.utils import load_model
import mlx.core as mx, numpy as np, soundfile as sf, librosa

model = load_model("8bit")              # a downloaded folder
ref = mx.array(librosa.load("voice.wav", sr=model.sample_rate)[0])
res = list(model.generate(text="Hello from Moonbox.", ref_audio=ref, ref_text="Exact transcript of voice.wav."))
sf.write("out.wav", np.concatenate([np.array(r.audio) for r in res]), res[0].sample_rate)

8bit/, 6bit/ and 4bit/ quantize the transformer layers (MLX affine, group size 64); the codec is not quantized. License: Fish Audio Research License: research and non-commercial use only; commercial use needs a license from Fish Audio. See LICENSE.md and NOTICE.md (what was changed).

How they were checked. Voice cloning of 6 English and 4 Chinese sentences, two seeds each, with the original (fish-speech's own inference code, PyTorch bfloat16) and every MLX build, the original's sampling (temperature 0.7, top-p 0.7, top-k 30). Judged like the original: Whisper large-v3-turbo WER / CER, WavLM speaker similarity to the prompt, DNSMOS. Pass: error rate at most 2 points above the original's, similarity at most 0.03 below, DNSMOS at most 0.2 below.

Precision	Size	WER en / CER zh	Speaker similarity en / zh	DNSMOS en / zh	Real-time factor, M5 24 GB
original (PyTorch bf16, MPS)	-	0.0 % / 0.0 %	0.952 / 0.949	3.21 / 2.76	4.2
bf16/	11.9 GB	0.0 % / 0.0 %	0.956 / 0.952	3.24 / 2.92	~5
fp16/	11.9 GB	0.0 % / 0.0 %	0.964 / 0.951	3.21 / 2.86	4.7
8bit/	8.5 GB	0.0 % / 0.0 %	0.953 / 0.953	3.22 / 2.83	3.2
6bit/	7.6 GB	0.0 % / 0.0 %	0.966 / 0.945	3.24 / 2.80	3.1
4bit/	6.7 GB	0.0 % / 0.0 %	0.963 / 0.944	3.22 / 3.01	3.0

Real-time factor = generation time / audio length, one sentence at a time (slower than real time on a 24 GB M5).

How these were made

From Fish Audio S2 Pro at revision 1de9996b6b: the original weights loaded in MLX and stored as bf16 or fp16, or quantized with MLX's affine quantization (8bit, 6bit, 4bit, group size 64; layers MLX can't quantize, such as convolutions and norms, stay bf16). Every build was checked against the original before upload; the sha256 column above is what the files must hash to.

License

other (Fish Audio research license). These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.

---

Original model card

Fish Audio S2 Pro

Technical Report | GitHub | Playground

Fish Audio S2 Pro is a leading text-to-speech (TTS) model with fine-grained inline control of prosody and emotion. Trained on over 10M+ hours of audio data across 80+ languages, the system combines reinforcement learning alignment with a dual-autoregressive architecture. The release includes model weights, fine-tuning code, and an SGLang-based streaming inference engine.

Architecture

S2 Pro builds on a decoder-only transformer combined with an RVQ-based audio codec (10 codebooks, ~21 Hz frame rate) using a Dual-Autoregressive (Dual-AR) architecture:

• Slow AR (4B parameters): Operates along the time axis and predicts the primary semantic codebook.
• Fast AR (400M parameters): Generates the remaining 9 residual codebooks at each time step, reconstructing fine-grained acoustic detail.

This asymmetric design keeps inference efficient while preserving audio fidelity. Because the Dual-AR architecture is structurally isomorphic to standard autoregressive LLMs, it inherits all LLM-native serving optimizations from SGLang — including continuous batching, paged KV cache, CUDA graph replay, and RadixAttention-based prefix caching.

Fine-Grained Inline Control

S2 Pro enables localized control over speech generation by embedding natural-language instructions directly within the text using [tag] syntax. Rather than relying on a fixed set of predefined tags, S2 Pro accepts free-form textual descriptions — such as [whisper in small voice], [professional broadcast tone], or [pitch up] — allowing open-ended expression control at the word level.

Common tags (15,000+ unique tags supported):

[pause] [emphasis] [laughing] [inhale] [chuckle] [tsk] [singing] [excited] [laughing tone] [interrupting] [chuckling] [excited tone] [volume up] [echo] [angry] [low volume] [sigh] [low voice] [whisper] [screaming] [shouting] [loud] [surprised] [short pause] [exhale] [delight] [panting] [audience laughter] [with strong accent] [volume down] [clearing throat] [sad] [moaning] [shocked]

Supported Languages

S2 Pro supports 80+ languages.

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

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

Other supported languages: sv, it, tr, no, nl, cy, eu, ca, da, gl, ta, hu, fi, pl, et, hi, la, ur, th, vi, jw, bn, yo, xsl, 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, and more.

Production Streaming Performance

On a single NVIDIA H200 GPU:

• Real-Time Factor (RTF): 0.195
• Time-to-first-audio: ~100 ms
• Throughput: 3,000+ acoustic tokens/s while maintaining RTF below 0.5

Links

• Fish Speech GitHub
• Fish Audio Playground
• Blog & Tech Report

Technical Report

If you find our work useful, please consider citing our report:

@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},
}

License

This model is licensed under the Fish Audio Research License. Research and non-commercial use is permitted free of charge. Commercial use requires a separate license from Fish Audio — contact business@fish.audio.