Microsoft AI ships MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash

Why it matters

【Key points】

On Oct 1, Microsoft AI released MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, along with two voice models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash.

The transcription model covers 60 languages with continuous language detection and returns first partial results in just over 100 ms. Microsoft says it ranks No. 1 for final and partial accuracy on Artificial Analysis. Introductory price: $0.54 per audio hour through year-end.

MAI-Voice-2.1 supports 23 languages and 26 locales, keeping one voice across languages with native accents, at $22 per 1M characters. Flash targets high-volume, low-latency use with about 150 ms end-to-end latency at $15 per 1M characters. Both support voice cloning from a few seconds of audio, with consent guardrails.

All three are available via Microsoft Foundry, MAI Playground, Vercel and Azure Voice Live; the voice models are also on OpenRouter.

【Verifiable】

Announcement: https://microsoft.ai/news/our-first-streaming-transcription-model/

The "2x faster than the closest competitor" claim is from Microsoft's internal evaluation. X discussion peaked two days after launch (about 97 posts on Oct 3) and has faded since.

【Our read】

1. Microsoft now supplies both ends of the voice-agent loop, so Azure customers building phone agents no longer need a separate speech vendor, which pressures voice-only companies.

2. Fast partials let agents start reasoning and calling tools before the user finishes speaking.

3. Microsoft gives no turn-taking data. Low latency does not help if the agent replies at the wrong moment.