Microsoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis

Michal Sutter at MarkTechPost reports that Microsoft AI released MAI-Transcribe-2-Streaming on October 1, 2026, alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash text-to-speech. It’s the real-time sibling of batch MAI-Transcribe-2 from September: 60 languages with continuous auto-detect, audio in and text out while the speaker is still talking.

Artificial Analysis ranks it #1 of 38 on AA-WER Streaming for both final and first partial transcripts — 2.5% WER at 0.13s to final and 0.12s to first partial on a mix of AA-AgentTalk, VoxPopuli and Earnings22. Runners-up include Grok Voice Transcribe 2.0 (2.7% at 0.49s) and Muse Voice Transcribe (3.1% at 0.16s). Cartesia Ink-2 is faster to final (0.07s) at 4.0% WER.

Introductory price is $0.54 per hour of audio through end of 2026 — higher than xAI and Meta on the comparison table, roughly in line with Google’s estimated rate. Integration paths: OpenAI-compatible Realtime WebSocket and Azure Speech SDK. Public preview, no SLA, no open weights.

We’re for streaming STT that makes partials trustworthy. Voice agents don’t wait for the period; they start tool calls mid-sentence. Matching 2.5% WER on first partial and final is the headline that matters more than a pure speed crown Cartesia can still claim.

Useful AI here is accessibility and product fluency: live captions, dictation, hands-free agents. Pairing with MAI-Voice-2.1-Flash (45s of audio at 150ms end-to-end for $15 per 1M characters in the write-up) sketches a full Microsoft voice loop. That’s competitive pressure on everyone selling realtime voice stacks.

The downside if closed streaming STT wins is the usual: no weights, no SLA in preview, and pricing that sits above Meta and xAI while accuracy leads. Enterprises will pay for WER; startups may not. Continuous language detection across 60 languages is powerful and raises the familiar surveillance/recording questions any always-on mic product inherits.

I’m watching whether Microsoft keeps the #1 AA rank once rivals retrain, and whether LiveKit support and an SLA arrive before buyers treat this as production-critical.

Microsoft says internal tests show words appearing 2x faster than its closest competitor. The model sits on Artificial Analysis’s accuracy-versus-latency Pareto frontier in the MarkTechPost summary.