Microsoft AI unveils first streaming transcription model MAI-Transcribe-2-Streaming

Microsoft AI, the AI division of Microsoft, announced three models together, and the range of languages involved is telling. The streaming transcription model handles 60 languages including Japanese, while the multilingual text-to-speech model covers 23 languages, and the voice models can keep one voice while switching languages with native accents. The pricing is set per use: $0.54 per hour of audio, $22 per million characters, and $15 per million characters for the high-speed Flash version, which Microsoft says can generate 45 seconds of audio with 150 milliseconds of latency.

The release places Microsoft against other large players the article names as already moving into voice AI, including OpenAI's streaming transcription model, Google's live model, and Meta's open-source speech recognition system. Microsoft's answer is to bundle transcription and speech synthesis for building voice agents, and to offer them not only in Microsoft Foundry but also in MAI Playground and elsewhere, alongside a combined demo called Chatter. In that demo, speaking Japanese is recognized, but responses are not configured in Japanese.

The voice replication feature is the part with the clearest guardrails: Microsoft says it requires review and approval, and only voices with the person's consent can be synthesized. What all of this ultimately hinges on is how builders respond to the accuracy ranking Microsoft cites from Artificial Analysis and to the low-latency claims, balanced against the introductory pricing that runs only through the end of the year.