Artificial Analysis Word Error Rate Index（非流式）
Artificial Analysis Word Error Rate Index（非流式）
% of words transcribed incorrectly · Lower is better · AA-WER v2 incorporates 3 datasets: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%), Earnings22-Cleaned-AA (25%)
注意： 对于 Earnings22，如果模型因时间限制而无法可靠处理完整长度的音频，我们会将音频切分为约 9 分钟的片段（相关模型：GPT-4o Mini Transcribe, OpenAINova 2 Pro, AmazonGPT-4o Transcribe, OpenAI）。对于时间限制更短的模型，我们会将音频切分为约 30 秒的片段（相关模型：Inworld STT 1Canary Qwen 2.5B, NVIDIAQwen3 ASR Flash, Alibaba）。
AA-WER（非流式）按数据集
AA-WER（非流式）：AA-AgentTalk 数据集
% of words transcribed incorrectly on the AA-AgentTalk dataset · Lower is better
清洗数据集对比
VoxPopuli：公开数据的清洗子集 vs. 原始子集
% WER (word error rate) · Lower is better
排序依据
注意：清洗版本会从参考文本中移除转写错误，从而为模型评估提供更准确的基准真值。
API 基准测试
Artificial Analysis Word Error Rate Index（非流式）vs. 价格
% of words transcribed incorrectly · Lower is better · AA-WER v2 incorporates 3 datasets: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%), Earnings22-Cleaned-AA (25%) · USD per 1000 minutes of audio
Most attractive quadrant
速度因子
Input audio seconds transcribed per second · Higher is better
转写价格
USD per 1000 minutes of audio
关键指标摘要与更多信息
服务商 | 更多详情 | ||||
|---|---|---|---|---|---|
Qwen3.5 Omni Flash |  Alibaba Cloud | 13.5% | 75.9 | 0.00 | |
Qwen3.5 Omni Plus | Alibaba Cloud | 3.5% | 190.4 | 0.00 | |
Fun-Realtime-ASR-preview | Alibaba Cloud | 1.7% | 20.9 | 0.00 | |
Nova 2 Pro |  Amazon Bedrock | 4.9% | 23.9 | 3.10 | |
Amazon Transcribe | Amazon Bedrock | 4.1% | 30.4 | 6.00 | |
Universal-3 Pro |  AssemblyAI | 3.1% | 109.0 | 3.50 | |
Universal, AssemblyAI | AssemblyAI | 3.8% | 122.4 | 2.50 | |
MAI-Transcribe-2 |  Microsoft AI | 2.0% | 314.2 | 1.67 | |
MAI-Transcribe-1.5 | Microsoft AI | 2.4% | 80.3 | 6.00 | |
transcribe-03-2026 |  Cohere | 4.6% | 122.3 | 0.00 | |
Nova-3 |  Deepgram | 5.2% | 618.3 | 4.30 | |
Nova-3, Telnyx |  Telnyx | 4.8% | 258.2 | 7.40 | |
Nova-2, Telnyx | Telnyx | 5.1% | 465.1 | 7.40 | |
Scribe v2 |  ElevenLabs | 2.2% | 71.0 | 3.67 | |
Solaria-1, Gladia |  Gladia | 4.1% | 75.2 | 10.17 | |
Solaria-3, Gladia | Gladia | 3.2% | 70.2 | 10.16 | |
Chirp 3, Google |  Google | 4.3% | 30.2 | 16.00 | |
Chirp | Google | 31.2% | 13.8 | 16.00 | |
Gemini 3.5 Transcribe | Google | 2.6% | 90.6 | 5.00 | |
Gemini 3.1 Pro Preview (High) | Google | 2.8% | 7.4 | 18.15 | |
Gemini 3.1 Pro Preview (Low) | Google | 3.6% | 8.2 | 7.72 | |
Gemini 3 Flash (High) | Google | 2.9% | 20.3 | 13.70 | |
Gemini 2.5 Flash Lite | Google | 5.2% | 80.0 | 6.56 | |
Gemini 2.5 Flash | Google | 5.1% | 74.5 | 6.66 | |
Gemini 2.5 Pro | Google | 2.9% | 10.6 | 11.39 | |
Gemini 3.1 Flash-Lite Preview (Minimal) | Google | 3.4% | 85.1 | 5.83 | |
Gradium Speech-to-Text |  Gradium | 6.8% | 2.3 | 13.00 | |
Grok Voice Transcribe 2.0 |  SpaceXAI | 2.3% | 181.8 | 1.67 | |
Grok Voice Transcribe 1.0 | SpaceXAI | 4.0% | 164.9 | 1.67 | |
Inworld STT 1 |  Inworld | 3.9% | 150.9 | 2.50 | |
Voxtral Mini Transcribe 2 |  Mistral | 3.6% | 87.0 | 3.00 | |
Voxtral Small | Mistral | 2.8% | 66.6 | 4.00 | |
Voxtral Mini |  DeepInfra | 3.8% | 51.2 | 1.00 | |
Modulate STT Batch English VFast |  Modulate | 4.2% | 52.9 | 0.42 | |
Parakeet TDT 0.6B V3, Togetherai |  Together AI | 4.5% | 448.7 | 1.50 | |
Canary Qwen 2.5B, NVIDIA |  Replicate | 4.3% | 8.8 | 0.74 | |
Parakeet TDT 0.6B V2, NVIDIA |  NVIDIA | 6.4% | 100.8 | 0.00 | |
Parakeet RNNT 1.1B | Replicate | 5.4% | 6.3 | 1.91 | |
GPT Transcribe, OpenAI |  OpenAI | 3.3% | 56.8 | 4.50 | |
GPT-4o Transcribe | OpenAI | 4.0% | 36.9 | 6.00 | |
GPT-4o Mini Transcribe | OpenAI | 4.5% | 52.5 | 3.00 | |
Smallest AI Pulse Pro |  Smallest.ai | 2.4% | 285.2 | 4.00 | |
Resonant-1 |  Reson8 | 3.4% | 344.3 | 3.60 | |
Rev AI |  Rev AI | 5.9% | 13.0 | 3.33 | |
Smallest AI Pulse | Smallest.ai | 4.4% | 252.6 | 5.00 | |
Soniox v5 Async |  Soniox | 3.8% | 44.2 | 1.66 | |
Speechmatics Melia |  Speechmatics | 4.9% | 180.7 | 4.00 | |
Speechmatics Standard | Speechmatics | 5.1% | 102.7 | 7.50 | |
Speechmatics Enhanced | Speechmatics | 4.0% | 80.0 | 12.50 | |
StepAudio 2.5 ASR, StepFun |  StepFun | 4.7% | 110.4 | 0.37 | |
StepAudio 3 ASR, StepFun | StepFun | 1.7% | 70.4 | 6.67 | |
Whisper Large v3 Turbo |  Groq | 4.6% | 116.6 | 0.67 | |
Whisper Large v3 Turbo, Telnyx | Telnyx | 5.5% | 39.5 | 15.00 | |
Wizper Large v3 |  fal.ai | 4.7% | 257.2 | 0.50 | |
Incredibly Fast Whisper | Replicate | 5.7% | 56.2 | 1.49 | |
Whisper Large v3 | Replicate | 10.1% | 2.6 | 4.23 | |
Whisper Large v3 | fal.ai | 4.1% | 117.3 | 1.15 | |
Whisper Large v3 | Together AI | 4.5% | 350.8 | 1.50 | |
Whisper Large v2 | OpenAI | 4.1% | 28.2 | 6.00 | |
Zoom Scribe V2 Pro |  Zoom | 2.8% | 75.1 | 3.33 |
常见问题
StepAudio 3 ASR 以最低的 AA-WER（Artificial Analysis 词错误率）1.7% 在 58 个已评估模型中领先。
按准确度（AA-WER）排名的顶级语音转文本模型为：1. StepAudio 3 ASR（1.7%）、2. Fun-Realtime-ASR-preview（1.7%）、3. MAI-Transcribe-2（2.0%）、4. Scribe v2, ElevenLabs（2.2%）和5. Grok Voice Transcribe 2.0（2.3%）。AA-WER 越低，表示转写准确度越高。
Nova-3 最快，速度因子为实时的 618.3x，其次是 Nova-2, Telnyx（465.1x）和 Parakeet TDT 0.6B V3, Togetherai（448.7x）。速度因子越高，转写越快。
StepAudio 2.5 ASR 最实惠，价格为每 1,000 分钟 $0.3667，其次是 Modulate STT Batch English VFast（$0.417）和 Wizper (L, v3), fal.ai（$0.50）。
Voxtral Small, Mistral 是最准确的开放权重模型，AA-WER 为 2.8%。在总共 58 个已评估模型中，有 13 个开放权重模型。
按准确度排名的顶级开放权重语音转文本模型为：1. Voxtral Small, Mistral（AA-WER 2.8%）、2. Inkling (256K), Thinking Machines（AA-WER 3.5%）和3. Voxtral Mini Transcribe 2, Mistral（AA-WER 3.6%）。
最佳模型取决于你的优先级。使用散点图可视化准确度（AA-WER）、速度与价格之间的权衡。对于需要高准确度的应用，优先选择 AA-WER 更低的模型；对于实时应用，关注速度因子；对于成本敏感的工作负载，比较价格图表。