走样 Zouyang · A Chinese ASR Benchmark Beyond WER/CER

最好的中文语音转文字模型是什么?

中文 ASR 榜单要么是英文测试, 要么干净得像朗诵棚. WER (中文里是 CER) 只看每个字 对不对, 但每个错误在意思层面的影响可以非常不一样. 「我喜欢」和「我不喜欢」只差 1 个字, 意思整个反了. 走样把错误分成 4 大类: 意思翻转扣 35 分, 实体认错扣 10 分, 口语填充扣 1 分. 每个模型除了名次, 还有一份错误的具体分布.

Most Chinese-ASR leaderboards are English-only or studio-clean, and WER (CER in Chinese) just counts characters without asking whether the meaning held up. Zouyang sorts errors into 4 groups: a flipped negation costs 35 points, a mangled entity 10, a filler word 1. Every model gets a breakdown of where its errors fall, not just a rank.

31 models · 100 samples · 78.3 min audio · judge: deepseek_flash · rubric: v6.8, v6.9 · updated 2026-09-23
Fidelity = 100 x (1 - deductions-per-100-chars / 20); higher is better.

榜单 · Leaderboard (专用 ASR / dedicated ASR)

# 模型 Model 厂商 Vendor 类型 保真分Fidelity 意思翻转Meaning flips实体认错Entities文化语境Cultural转写保真Fidelity 单价 Price
1MAI-Transcribe-2mai_transcribe_2MicrosoftAPI90.4-2.9-2.8-0.2-3.8$0.0017/min
2MiMo-V2.5-ASRmimoXiaomiOSS87.5-1.4-4.7-0.5-5.8$0.0012/min
3Qwen3-ASR-FlashflashAlibabaAPI84.5-3.5-4.9-0.9-6.2$0.0021/min
4Qwen-Audio-3.1-ASRqwen_audio_3_1_asrAlibabaAPI84.4-4.6-4.0-0.9-6.1-
5Scribe v2elevenlabsElevenLabsAPI83.9-3.2-4.1-2.0-6.8$0.0037/min
6Doubao 2.0doubaoByteDanceAPI82.1-4.8-6.4-0.5-6.3-
7Confucius4-R2T2r2t2NetEase YoudaoOSS81.9-4.2-5.9-1.1-7.0-
8Soniox stt-async-v5sonioxSonioxAPI81.0-0.9-6.2-2.9-9.0$0.0017/min
9GPT Transcribegpt_transcribeOpenAIAPI80.9-2.9-4.3-1.2-10.7$0.0045/min
10VibeVoice-ASRvibevoiceMicrosoftOSS80.4-5.3-5.7-1.2-7.4-
11GPT-4o-mini-transcribewhisperOpenAIAPI76.6-3.8-6.5-3.3-9.8$0.003/min
12Qwen3-ASR-1.7Bqwen3_ossAlibaba OSSOSS76.1-5.4-7.6-1.9-9.0-
13Gemini 3.5 Transcribegemini_3_5_transcribeGoogleAPI74.3-4.5-6.1-2.1-12.9$0.005/min
14GLM-ASR-Nanoglm_asr_nanoZhipu OSSOSS73.1-7.8-7.2-1.1-8.6-
15Step-Audio 2.5 ASRstepfunStepFunAPI71.3-8.4-9.1-1.1-10.1-
16FireRedASR2-AEDfirered_aedXiaohongshu OSSOSS69.3-4.3-14.2-2.1-10.1自建 self-host
17Fun-ASR-Nanofunasr_nanoAlibaba/DAMO OSSOSS63.5-8.3-7.6-1.1-8.3-
18Whisper-Large-v3-turbowhisper_v3_turboOpenAI OSSOSS58.8-8.2-11.3-5.6-16.1自建 self-host
19Apple SpeechAnalyzer (macOS 26)apple_speechanalyzerAppleLOCAL48.6-4.9-23.9-6.5-16.1自建 self-host
20FireRedASR2-CTCfireredXiaohongshu OSSOSS35.7-4.4-31.8-6.8-21.2自建 self-host

页面默认只显示基准线 (Qwen3-ASR-1.7B) 以上的 12 个模型, 其余行在这里 一并列出, 顺序和分数与交互版一致. The live board hides rows below the qwen3_oss baseline by default; all of them are listed here.

多模态 LLM 转写 · LLM-as-ASR (reference)

# 模型 Model 厂商 Vendor 类型 保真分Fidelity 意思翻转Meaning flips实体认错Entities文化语境Cultural转写保真Fidelity 单价 Price
1Gemini 3.8 Flash (LLM-as-ASR · thinkingLevel=low)gemini_3_8_flash_sttGoogleAPI84.2-3.7-1.9-0.9-9.3-
2Gemini 3 Pro (LLM-as-ASR · thinkingLevel=low)gemini_pro_sttGoogleAPI82.9-3.1-3.7-0.2-10.1-
3MiMo V2.5 (LLM-as-ASR)mimo_sttXiaomiAPI79.6-5.0-4.5-1.2-9.8-
4Gemini 3.7 Flash (LLM-as-ASR · thinkingLevel=low)gemini_3_7_flash_sttGoogleAPI78.9-5.7-2.6-1.3-11.6-
5Gemini 3 Flash (LLM-as-ASR)gemini_flash_sttGoogleAPI76.6-5.0-3.7-1.3-13.4-
6Gemini 3.5 Flash (LLM-as-ASR)gemini_3_5_flash_sttGoogleAPI76.4-3.4-4.7-2.5-13.0-
7Gemini 3.6 Flash (LLM-as-ASR)gemini_3_6_flash_sttGoogleAPI69.9-6.1-4.7-2.3-17.1-
8Gemini 3.1 Flash Lite (LLM-as-ASR)gemini_sttGoogleAPI63.8-5.6-9.4-3.7-17.5-
9Voxtral-Mini-4B-RealtimevoxtralMistral OSSOSS55.7-11.2-9.5-5.4-18.3-
10MOSS-Audio-8B-Instruct (LLM-as-ASR)moss_audioOpenMOSS / MOSI.AI OSSOSS46.0-21.5-7.8-2.7-19.8-
11Gemini 3.5 Flash-Lite (LLM-as-ASR)gemini_3_5_flash_lite_sttGoogleAPI47.8-10.9-10.0-4.8-26.4-

多模态 LLM 不是专用 ASR, 默认不进主榜, 单独作为参照. Multimodal LLMs are excluded from the main board by default and shown here for reference only.

四个维度 · The four groups

意思翻转 · Meaning flips — polarity, semantic-inversion, hallucination, numeric

实体认错 · Entities — brand, tech, proper-noun

文化语境 · Cultural — culture, dialect

转写保真 · Fidelity — mishear-severe, content-drop, paraphrase, mishear-minor, filler, format

数据 · Data

完整聚合数据: eval_aggregate.json (scores, per-category deductions, pricing, per-sample results). 不含音频与参考转写文本. No audio or reference transcripts are published. llms.txt