whatmodel.sh / tts / Fish Audio S2 Pro
Fish Audio S2 Pro
↗ share
+ contribute
Fish Audio S2 Pro is an open-weights text-to-speech model from Fish Audio, released March 2026. Its published size is 4.56B parameters. Licence: fish-audio-research-license.
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Fish Audio S2 Pro, out of 148 with a score, then more from Fish Audio.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 35 |  Sonic 3.5 | 63±2 | comparison only | ||
| 36 |  MiniMax Speech-02-HD | 63±3 | api $100.00/M chars | ||
| 37 | Fish Audio S2 Pro | 63±1 | api $15.00/M chars +local | ||
| 38 |  Eleven Flash v2.5 | 63±2 | api $50.00/M chars | ||
| 39 |  IndexTTS (v1) | 62±3 | local | ||
| more from Fish Audio | |||||
| 5 | OpenAudio S1 | 71±8 | api $15.00/M chars | ||
| 31 | Fish Audio S2.1 Pro | 63±3 | api $15.00/M chars | ||
| 88 | Fish Speech 1.2 | 49±14 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #15 of 37 · 50% of its score · 8 tests
Beats Gemini 3.1 Flash TTS Preview, Fun-CosyVoice3-0.5B-2512, IndexTTS-2 here, though they rank higher overall. Trails F5-TTS v1 Base, IndexTTS (v1) here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) +1 | 64.9% ✎ | █·█·██·▄·▄·▄ beats 0/10 | 14% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) | 72.8% ✎ | ▃··▃█▅▅··▃·▃ beats 0/10 | 9.2% |
| Seed-TTS-Eval · test-en/test-zh macro mean error | 1.3% ✎ | ▄··▄▄····█▄▄ beats 3/6 | 6.1% |
| MiniMax multilingual test set · 21-language mean error (UltraEval-Audio common set) | 8.1% ✎ | █··█·····█·█ beats 0/3 | 5.0% |
| Seed-TTS test-en · Whisper large-v3 WER +1 | 1.8% ✎ | ▃··▃▅··▅·█▃▅ beats 6/11 | 5.0% |
| Seed-TTS test-en WER (transcriber not stated) | 1.6% ✎ | ▃··········█ beats 2/3 | 5.0% |
| Seed-TTS test-zh · Paraformer CER +1 | 1.0% ✎ | ▂··▄··▂▂█·▄▂ beats 7/11 | 4.3% |
| Seed-TTS test-en · UTMOS (predicted MOS) | 2.80 / 5 ✎ | █·····██···█ beats 3/3 | 2.0% |
other tests16 not in a task · 49% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| ZTTS1-Eval in-the-wild · 10-language mean WER | 2.9% ✎ | ▄·····█··▄██ beats 3/7 | 8.2% |
| ZTTS1-Eval Clean (FLEURS-R) · 9-language mean WER | 4.5% ✎ | ▄·····▄█▄█·▄ beats 2/7 | 7.1% |
| ZTTS1-Eval in-the-wild · 10-language mean MSR-UTMOS | 2.71 / 5 ✎ | █··▄█▄····▄▄ beats 2/7 | 6.9% |
| ZTTS1-Eval Clean (FLEURS-R) · 9-language mean MSR-UTMOS | 3.21 / 5 ✎ | ▃▃█···▃···▃▃ beats 4/7 | 6.1% |
| RW-Voice-EQ TTS · Acoustic Quality (human) | 3.34 / 5 ✎ | █·█········▄ beats 3/4 | 5.0% |
| RW-Voice-EQ TTS · Language Code-Switching (human) | 3.48 / 5 ✎ | █··▄·······▄ beats 3/3 | 3.4% |
| ZTTS1-Eval · FLEURS-R English TTSDS2 prosody | 87.5% ✎ | █··███·····█ beats 2/4 | 2.9% |
| RW-Voice-EQ TTS · Voice Identity (human) | 3.53 / 5 ✎ | █·····▄····█ beats 3/4 | 2.9% |
8 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity | 54.0% ✎ | ▃·····▂▂▂▃█▆ beats 5/15 | 2.6% |
| MiniMax multilingual test set · 23-language mean error | 5.1% ✎ | ▃·▃·▃·····▃█ beats 3/6 | 1.3% |
| Arabic TTS Arena · community blind A/B rating, 2026-09-24 snapshot | 1057.4 ✎ | ▂▂··▃·█▅···▂ beats 8/12 | 1.3% |
| CV3-Eval multilingual voice clone · 9-language mean error | 4.6% ✎ | ▃·····▃···▅█ beats 5/6 | 1.0% |
| RW-Voice-EQ TTS · Long-Form Stability (human) | 3.50 / 5 ✎ | █······█·███ beats 1/4 | 0.4% |
| RVCBench LibriTTS clean voice cloning · Whisper-medium WER | 4.0% ✎ | ▂▂▂··▂·▃▃▃█▃ beats 15/15 | 0.0% |
| Long-TTS-Eval · Fish truncated English WER lab | 4.4% ✎ | █··········█ beats 1/1 | — |
| Long-TTS-Eval · Fish truncated Chinese CER lab | 5.9% ✎ | █··········█ beats 1/1 | — |
$ where to run it
Through an API
api · $15.00/M chars · docs.fish.audio ↗ · price sourceFish Audio lists s2-pro at $15.00 per 1M UTF-8 bytes of input text. Stored as per 1M characters, which equals bytes only for ASCII (e.g. English) text; non-ASCII characters take 2-4 bytes, so per-character cost is higher for those scripts. Product: Fish Audio TTS API. API model: s2-pro. Settings: pay-as-you-go. Price basis: published list rate (byte-billed). Exclusions: none stated. Currency: USD. Fish Audio official pricing docs, observed 2026-09-25.
Other ways to run it
version
- Run it as published · needs ~11.44 GB, NVIDIA only · from github.com/fishaudio (official)
git clone https://github.com/fishaudio/fish-speech && cd fish-speech && pip install -e . && hf download fishaudio/s2-pro --local-dir checkpoints/s2-pro && python fish_speech/models/text2semantic/inference.py --text 'Hello world' && python fish_speech/models/dac/inference.py -i output/codes_0.npy --checkpoint-path checkpoints/s2-pro/codec.pth
Official weights and documentation: huggingface.co/fishaudio ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 30%, ZONOS2 Technical Report (ZTTS… 28%, RW-Voice-EQ Bench (Hume AI) 12% and 6 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 UltraEval-Audio (OpenBMB) · TTS replication results | 5 | 30% |
 ZONOS2 Technical Report (ZTTS1-Eval results) | 4 | 28% |
| RW-Voice-EQ Bench (Hume AI) | 4 | 12% |
 nineninesix/gepard-1.0 model card (Seed-TTS-eval benchmark table) | 3 | 11% |
| Higgs Audio v3 comparison tables in SGLang-Omni cookbook (Boson AI contribution) | 3 | 8% |
| ZTTS1-Eval results snapshot (Zyphra) | 1 | 3% |
| RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2) | 2 | 3% |
| Fish Audio S2 Technical Report · lab-reported | 2 | 3% |
| Arabic TTS Arena (Navid AI) · leaderboard snapshot 2026-09-24 | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 62.6 |
not counted (2)
| Fish Audio S2 Technical Report | Long-TTS-Eval · Fish truncated English WER [correct] | 4.4% | fewer than three models have this test yet, or this source ran it for this model alone |
| Fish Audio S2 Technical Report | Long-TTS-Eval · Fish truncated Chinese CER [correct] | 5.9% | fewer than three models have this test yet, or this source ran it for this model alone |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json