Microsoft Ships Its Own Voice Stack: 60-Language Streaming Transcription + Seconds-Level Voice Cloning

What Happened

On October 1, Microsoft AI unveiled three homegrown voice models on its official blog microsoft.ai: streaming transcription model MAI-Transcribe-2-Streaming, plus two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-Flash, covering both speech recognition and speech synthesis.

The transcription model is all about real time: 60 languages, continuous automatic language detection, ~100ms first-word latency, priced at $0.54 per audio hour (promotional price this year). Microsoft claims it ranks #1 on Artificial Analysis for both final and partial transcription accuracy. On the TTS side, MAI-Voice-2.1 covers 23 languages and 26 regions, with same-speaker cross-language support that keeps local accents ($22 per million characters); the Flash edition pushes latency down to 150ms ($15 per million characters). Voice cloning needs only seconds of reference audio, but requires approval and ships with consent safeguards.

Key Facts

1. When & where: October 1, 2026, first published on the Microsoft AI official blog (microsoft.ai); covered consistently by The Decoder, RuntimeWire, TestingCatalog, and WindowsReport.
2. Transcription: MAI-Transcribe-2-Streaming: 60 languages, continuous automatic language detection, ~100ms first word; Microsoft claims #1 on Artificial Analysis final/partial transcription accuracy; $0.54/audio hour (promotional price this year).
3. Speech synthesis: MAI-Voice-2.1: 23 languages / 26 regions, same-speaker cross-language with local accents, $22 per million characters; Flash edition: 150ms latency, $15 per million characters.
4. Voice cloning: Clones a voice from seconds of reference audio; requires approval and includes consent safeguards.

Context

OpenAI had already launched its Realtime API and TTS models. Microsoft's three new models are all built in-house, with no outside models.