微软发布 MAI-Transcribe-2-Streaming 流式语音转文字模型，登顶 Artificial Analysis 榜单
Microsoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis - MarkTechPost
微软 AI 发布首款流式语音转文字模型 MAI-Transcribe-2-Streaming，支持 60 种语言并自动连续检测语种，在收到音频后 100 毫秒出头即输出首个部分转写结果。Artificial Analysis 的 AA-WER Streaming 测试显示，其最终转写与首个部分转写均为 2.5% WER，在 38 个模型中均排第一，最终结果延迟 0.13 秒。该模型定价为每小时音频 0.54 美元（2026 年底前的尝鲜价），可通过 Realtime API 或 Azure Speech SDK 接入，目前为公开预览，无 SLA 且未开放权重。
正文
当前显示原文，中文全文译文尚未就绪。
Microsoft AI has released MAI-Transcribe-2-Streaming, its first streaming speech-to-text (STT) model. It launched on October 1, 2026, alongside 2 text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Artificial Analysis ranks it #1 of 38 models for final and first partial transcript accuracy. The model targets voice agents, live captions and dictation, where latency decides the experience.
What Microsoft Shipped
MAI-Transcribe-2-Streaming is the real-time sibling of the batch MAI-Transcribe-2, released in September. It transcribes 60 languages with automatic, continuous language detection. Audio streams in continuously, and text streams back while the speaker is still talking.
The model emits its first hypotheses, called partials, just over 100ms after receiving audio. It revises those partials as context arrives, then commits a stable final transcript. An agent can therefore start reasoning or calling tools mid-sentence. Microsoft team states its internal tests show words appearing 2x faster than its closest competitor.
What Artificial Analysis Measured
The AA-WER Streaming index uses about 8 hours of audio. The mix is AA-AgentTalk (50%), VoxPopuli (25%) and Earnings22 (25%). Latency is timed from the end of speech, as detected by SileroVAD.
Final transcript: 2.5% WER at 0.13s after end of speech, #1 of 38 models; First partial transcript: 2.5% WER at 0.12s after end of speech, also #1; Runners-up: Grok Voice Transcribe 2.0 at 2.7% and 0.49s. Muse Voice Transcribe at 3.1% and 0.16s; Not the fastest: Cartesia Ink-2 (external endpoints) returns finals in 0.07s, but at 4.0% WER
The first partial is as accurate as the final transcript. That matters for agents that act before the speaker finishes. Microsoft also places the model on the accuracy versus latency Pareto frontier.
MAI-Transcribe-2-Streaming costs $0.54 per hour of audio. This is an introductory price through the end of 2026. Artificial Analysis normalizes it to $9.00 per 1,000 minutes. Batch MAI-Transcribe-2 costs $0.10 per hour. On streaming, Microsoft charges more than xAI and Meta, and roughly matches Google’s estimated rate.
How Developers Integrate It
Microsoft documents 2 integration paths. The Realtime API suits apps already using an OpenAI Realtime-compatible WebSocket. The Azure Speech SDK handles connection management, retries and audio streaming. Both return intermediate and final results.
The model is also available in the MAI Playground, through Vercel and Azure Voice Live. LiveKit support is listed as coming soon.
Microsoft pairs it with MAI-Voice-2.1-Flash for full voice loops. Flash generates 45s of audio at 150ms end-to-end latency for $15 per 1M characters. MAI-Voice-2.1 covers 23 languages and 26 locales at $22 per 1M characters.
Sources: Artificial Analysis, 9to5Mac (Muse pricing), DataCamp (Grok pricing), The Batch (Gemini pricing). Verified October 2, 2026.
Key Takeaways
#1 of 38 on AA-WER Streaming: 2.5% WER at 0.13s to final; First partials score the same 2.5% WER, at 0.12s; 60 languages with continuous automatic language detection; $0.54 per hour introductory price, higher than xAI and Meta; Public preview with no SLA and no open weights