Fish Speech S2 Pro 评测结果 / Evaluation Results
April 20, 2026 · View on GitHub
模型 / Model: fishaudio-s2-pro 评测日期 / Evaluation Date: 2026/04 Paper/Repo: fishaudio/s2-pro
指标说明 / Metrics:
- WER⬇️: Word Error Rate — 词错误率，越低越好 / lower is better
- CER⬇️: Character Error Rate — 字符错误率，越低越好 / lower is better
- SIM⬆️: Speaker Similarity — 说话人相似度，越高越好 / higher is better
Seed-TTS-Eval Benchmark
| task | dataset | WER/CER⬇️ | SIM⬆️ | eval_cli | note |
|---|---|---|---|---|---|
| tts | seed_tts_eval_en | 1.82 | 64.94 | [1] | |
| tts | seed_tts_eval_zh | 1.03 | 72.84 | [2] |
说明: 上述 Seed-TTS-Eval 结果（EN WER=1.82, ZH CER=1.03）与官方 README / 技术报告中公布的数值（EN=0.99, ZH=0.54）存在较大差距。原因是论文中公布的数值使用的是 fishaudio 线上服务的模型，而非开源模型；本评测基于公开权重
fishaudio/s2-pro及fish-speech仓库公开的本地推理路径完成，因此无法与论文数值对齐。相关讨论参考 fishaudio/fish-speech#1268。Note (EN): The Seed-TTS-Eval numbers above (EN WER=1.82, ZH CER=1.03) differ noticeably from the values reported in the official README / technical report (EN=0.99, ZH=0.54). The reason is that the paper's reported numbers were produced by fishaudio's online service model, not the open-source release. This evaluation is based on the publicly released weights
fishaudio/s2-proand the local inference pipeline provided in thefish-speechrepository, so the results are not directly comparable to the paper. See the related discussion in fishaudio/fish-speech#1268.
MiniMax TTS 多语言 Benchmark / MiniMax TTS Multilingual Benchmark
来源数据集 / Source dataset: MiniMaxAI/TTS-Multilingual-Test-Set
| task | language | dataset | WER/CER⬇️ | SIM⬆️ | eval_cli |
|---|---|---|---|---|---|
| tts | Arabic | minimax_tts_arabic | 8.02 (WER) | 73.26 | [3] |
| tts | Cantonese | minimax_tts_cantonese | 46.15 (CER) | 74.75 | [4] |
| tts | Chinese | minimax_tts_chinese | 1.08 (CER) | 76.90 | [5] |
| tts | Czech | minimax_tts_czech | 6.57 (WER) | 76.68 | [6] |
| tts | Dutch | minimax_tts_dutch | 1.40 (WER) | 71.59 | [7] |
| tts | English | minimax_tts_english | 2.39 (WER) | 78.71 | [8] |
| tts | Finnish | minimax_tts_finnish | 7.69 (WER) | 80.43 | [9] |
| tts | French | minimax_tts_french | 4.18 (WER) | 68.12 | [10] |
| tts | German | minimax_tts_german | 1.04 (WER) | 71.04 | [11] |
| tts | Greek | minimax_tts_greek | 12.45 (WER) | 79.71 | [12] |
| tts | Hindi | minimax_tts_hindi | 22.38 (WER) | 78.84 | [13] |
| tts | Indonesian | minimax_tts_indonesian | 2.82 (WER) | 73.29 | [14] |
| tts | Japanese | minimax_tts_japanese | 3.65 (CER) | 75.99 | [15] |
| tts | Korean | minimax_tts_korean | 1.59 (CER) | 75.32 | [16] |
| tts | Polish | minimax_tts_polish | 2.41 (WER) | 79.02 | [17] |
| tts | Portuguese | minimax_tts_portuguese | 1.53 (WER) | 79.31 | [18] |
| tts | Romanian | minimax_tts_romanian | 15.83 (WER) | 75.55 | [19] |
| tts | Russian | minimax_tts_russian | 3.89 (WER) | 76.31 | [20] |
| tts | Spanish | minimax_tts_spanish | 1.56 (WER) | 73.80 | [21] |
| tts | Thai | minimax_tts_thai | 7.59 (CER) | 74.87 | [22] |
| tts | Turkish | minimax_tts_turkish | 2.03 (WER) | 77.99 | [23] |
| tts | Vietnamese | minimax_tts_vietnamese | 15.74 (WER) | 70.87 | [24] |
Evaluation Commands
[1] python audio_evals/main.py --dataset seed_tts_eval_en --model fishaudio-s2-pro
[2] python audio_evals/main.py --dataset seed_tts_eval_zh --model fishaudio-s2-pro
[3] python audio_evals/main.py --dataset minimax_tts_arabic --model fishaudio-s2-pro
[4] python audio_evals/main.py --dataset minimax_tts_cantonese --model fishaudio-s2-pro
[5] python audio_evals/main.py --dataset minimax_tts_chinese --model fishaudio-s2-pro
[6] python audio_evals/main.py --dataset minimax_tts_czech --model fishaudio-s2-pro
[7] python audio_evals/main.py --dataset minimax_tts_dutch --model fishaudio-s2-pro
[8] python audio_evals/main.py --dataset minimax_tts_english --model fishaudio-s2-pro
[9] python audio_evals/main.py --dataset minimax_tts_finnish --model fishaudio-s2-pro
[10] python audio_evals/main.py --dataset minimax_tts_french --model fishaudio-s2-pro
[11] python audio_evals/main.py --dataset minimax_tts_german --model fishaudio-s2-pro
[12] python audio_evals/main.py --dataset minimax_tts_greek --model fishaudio-s2-pro
[13] python audio_evals/main.py --dataset minimax_tts_hindi --model fishaudio-s2-pro
[14] python audio_evals/main.py --dataset minimax_tts_indonesian --model fishaudio-s2-pro
[15] python audio_evals/main.py --dataset minimax_tts_japanese --model fishaudio-s2-pro
[16] python audio_evals/main.py --dataset minimax_tts_korean --model fishaudio-s2-pro
[17] python audio_evals/main.py --dataset minimax_tts_polish --model fishaudio-s2-pro
[18] python audio_evals/main.py --dataset minimax_tts_portuguese --model fishaudio-s2-pro
[19] python audio_evals/main.py --dataset minimax_tts_romanian --model fishaudio-s2-pro
[20] python audio_evals/main.py --dataset minimax_tts_russian --model fishaudio-s2-pro
[21] python audio_evals/main.py --dataset minimax_tts_spanish --model fishaudio-s2-pro
[22] python audio_evals/main.py --dataset minimax_tts_thai --model fishaudio-s2-pro
[23] python audio_evals/main.py --dataset minimax_tts_turkish --model fishaudio-s2-pro
[24] python audio_evals/main.py --dataset minimax_tts_vietnamese --model fishaudio-s2-pro