AI
Microsoft Launches MAI-Transcribe-2-Streaming — Its First Real-Time Transcription Model With Record Accuracy in 60 Languages
Microsoft AI announced MAI-Transcribe-2-Streaming on October 1, 2026, its first model built for real-time (streaming) speech transcription, alongside two new text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The transcription model supports 60 languages with automatic, continuous language detection that requires no language to be set in advance, producing its first partial hypothesis roughly 100 milliseconds after receiving audio and refining it as more context arrives before committing a final transcript.
In an independent evaluation of 38 models by Artificial Analysis, MAI-Transcribe-2-Streaming posted the most accurate final transcript with a 2.5% word error rate and a final-text latency of about 0.13 seconds after speech ends, ranking first among every model tested. Microsoft is pricing the model at an introductory $0.54 per hour, or $9 per 1,000 minutes, through the end of 2026, compared with $0.10 per hour for its non-streaming MAI-Transcribe-2 model.
MAI-Voice-2.1 converts text to speech in 23 languages and 26 locales while keeping a single voice's identity and native accent consistent across them, at roughly 550 milliseconds of latency and $22 per million characters. Its MAI-Voice-2.1-Flash variant targets high-volume, latency-sensitive use, generating up to 45 seconds of audio in about 150 milliseconds at $15 per million characters. Both voice models support cloning a voice from a short reference sample, with built-in consent safeguards Microsoft says are meant to prevent misuse, and all three models are now available through Microsoft Foundry.
Source: Microsoft AI
Need filming at the level you just read about? I reply within 24 hours.
Start your project