Microsoft AI's MAI-Transcribe-2-Streaming tops accuracy, 100ms latency

Microsoft AI is pairing a transcription engine with two text-to-speech models in a single voice-agent lineup. The transcription model's headline claim is speed and accuracy: Microsoft says it ranks first for accuracy on Artificial Analysis, covers 60 languages, and returns first partial results in just over 100 milliseconds, which Microsoft says lets voice agents respond while someone is still mid-sentence. Pricing is an introductory $0.54 per hour of audio through the end of the year.

The two speech models extend the same push. MAI-Voice-2.1 is designed to speak 23 languages in the same voice with a native accent in each, and the MAI-Voice-2.1-Flash variant is listed at 150 milliseconds of latency and $15 per million characters instead of $22. Both voice models can clone a voice from a few seconds of reference audio, and Microsoft says built-in safeguards are meant to prevent misuse. Distribution runs through Microsoft Foundry and the MAI Playground, and the two voice models are also on OpenRouter.

The adoption question hinges on trust as much as latency. In one test, about half of 4,000 participants thought the voices belonged to a real person — a result that cuts both ways for products that depend on voice agents sounding human, and for the safeguards Microsoft says are meant to keep cloning from being misused. How regulators and buyers react to that trade-off is likely to shape how quickly the models move from playground demos into live customer-facing systems.