The real-time transcription model produces its first hypotheses in a hundred milliseconds in 60 languages; the Flash variant of text-to-speech targets low-latency agents.
Microsoft AI has introduced its first streaming transcription model, MAI-Transcribe-2-Streaming. It handles 60 languages with automatic and continuous detection of spoken language, and ranks first in accuracy, for both final and draft transcriptions, on the Artificial Analysis benchmark. Its first text hypotheses arrive a little over 100 ms after receiving the audio, then are corrected as the context is enriched, allowing a voice agent to start reasoning or calling tools before the sentence is complete. The introductory rate is set at $0.54 per hour of audio until the end of the year.
One voice in 23 languages
Microsoft is also announcing MAI-Voice-2.1, its most advanced multilingual text-to-speech model, which covers 23 languages and 26 regional variants while keeping the same voice, with a native accent in each language. The MAI-Voice-2.1-Flash version, designed for high volumes, generates 45 seconds of audio with 150 ms of end-to-end latency, for $15 per million characters. Both models allow you to clone a voice from a few seconds of sample, with built-in consent guardrails.
Read the full article on the original source
Go to the source →