Microsoft AI 发布 MAI-Transcribe-2-Streaming 流式语音转文本模型，登顶 Artificial Analysis 实时榜单

微软 AI 发布了 MAI-Transcribe-2-Streaming：https://microsoft.ai/news/our-first-streaming-transcription-model/，这是其首个流式语音转文本（STT）模型。它于 2026 年 10 月 1 日推出，同时发布了两款文本转语音模型，MAI-Voice-2.1 和 MAI-Voice-2.1-Flash。Artificial Analysis：https://x.com/ArtificialAnlys/status/2105694108736188894 将其在最终和首次部分转录准确率方面排名 38 个模型中的第 1 名。该模型面向语音助手、实时字幕和语音记录，其中延迟决定体验。

MAI-Transcribe-2-Streaming 是批处理版本 MAI-Transcribe-2（于九月发布）的实时版。它支持 60 种语言的转录，并自动进行连续语言检测。音频连续输入，文本在说话者仍在讲话时即可流回。

该模型在接收音频约 100 毫秒后就会输出首个假设，即部分转录。随着上下文到来，它会修正这些部分转录，然后生成稳定的最终转录。因此，语音助手可以在句子未结束时就开始推理或调用工具。微软团队表示，其内部测试显示词语出现速度比最接近的竞争对手快 2 倍。

AA-WER 流式索引：https://artificialanalysis.ai/speech-to-text/streaming 使用了约 8 小时的音频。音频混合来自 AA-AgentTalk（50%）、VoxPopuli（25%）和 Earnings22（25%）。延迟从 SileroVAD 检测到的语音结束开始计时。

首个部分转录与最终转录一样准确。这对在说话者未说完前就采取行动的语音助手非常重要。微软还将该模型放在准确率与延迟的帕累托前沿上。

MAI-Transcribe-2-Streaming 的费用为每小时音频 0.54 美元。这是 2026 年底前的推广价格。Artificial Analysis 将其标准化为每 1000 分钟 9.00 美元。批处理版 MAI-Transcribe-2 的费用为每小时 0.10 美元。在流式服务上，微软收费高于 xAI 和 Meta，与 Google 的预估费用大致相当。

微软提供了两条集成路径。实时 API 适合已使用 OpenAI Realtime 兼容 WebSocket 的应用。Azure Speech SDK 负责连接管理、重试和音频流传输。两者都返回中间和最终结果。

该模型还可以在 MAI Playground 上使用：http://playground.microsoft.ai/，通过 Vercel：https://aka.ms/mai-vercel 以及 Azure Voice Live 使用。LiveKit 支持预计即将推出。

微软将其与 MAI-Voice-2.1-Flash：https://learn.microsoft.com/en-us/azure/ai-services/speech-service/mai-voices 配对，以实现完整语音循环。Flash 可生成 45 秒音频，端到端延迟为 150 毫秒，每 100 万字符收费 15 美元。MAI-Voice-2.1 涵盖 23 种语言和 26 个地区，每 100 万字符收费 22 美元。

来源：Artificial Analysis：https://artificialanalysis.ai/speech-to-text/streaming，9to5Mac（Muse 定价）：https://9to5mac.com/?p=1066961，DataCamp（Grok 定价）：https://www.datacamp.com/blog/grok-voice-transcribe-2-0，The Batch（Gemini 定价）：https://www.deeplearning.ai/the-batch/transcription-battles-heat-up。信息已于 2026 年 10 月 2 日验证。

需要与我们合作以推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等？请与我们联系：https://forms.gle/MJjjVDPS7whH8Ngs6

Datalab 推出 OmniExtractBench，以修正提取基准中的偏差和不透明性：https://pxllnk.co/hxplrq

以从业者为先的 AI/ML 新闻和分析，每月被 100 万+ 开发者和研究人员阅读。

Microsoft AI has released MAI-Transcribe-2-Streaming：https://microsoft.ai/news/our-first-streaming-transcription-model/, 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：https://x.com/ArtificialAnlys/status/2105694108736188894 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.

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.

The AA-WER Streaming index：https://artificialanalysis.ai/speech-to-text/streaming 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.

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.

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：http://playground.microsoft.ai/, through Vercel：https://aka.ms/mai-vercel and Azure Voice Live. LiveKit support is listed as coming soon.

Microsoft pairs it with MAI-Voice-2.1-Flash：https://learn.microsoft.com/en-us/azure/ai-services/speech-service/mai-voices 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：https://artificialanalysis.ai/speech-to-text/streaming, 9to5Mac (Muse pricing)：https://9to5mac.com/?p=1066961, DataCamp (Grok pricing)：https://www.datacamp.com/blog/grok-voice-transcribe-2-0, The Batch (Gemini pricing)：https://www.deeplearning.ai/the-batch/transcription-battles-heat-up. Verified October 2, 2026.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/MJjjVDPS7whH8Ngs6

Datalab Introduces OmniExtractBench to Fix Bias and Opacity in Extraction Benchmarks ：https://pxllnk.co/hxplrq

Practitioner-first AI/ML news and analysis, read by 1M+ developers and researchers every month.