Microsoft introduces real-time transcription and multilingual voice models

Microsoft has launched MAI-Transcribe-2-Streaming, a speech-to-text model designed to produce live transcripts with very low delay. Microsoft AI announces the models in an official blog post, alongside the new text-to-speech systems MAI-Voice-2.1 and MAI-Voice-2.1-Flash.

The company says MAI-Transcribe-2-Streaming supports 60 languages and can continuously identify the language being spoken. Instead of waiting until a speaker finishes a sentence, the model returns an initial text prediction in just over 100 milliseconds, according to Microsoft. It then updates that text as more audio arrives and finalizes stable wording when enough context is available.

This approach could help customer service tools, live captioning systems, and voice-controlled applications respond while a person is still speaking. Microsoft claims that, in its internal tests, words appear twice as quickly as with its closest competitor for real-time dictation and subtitles. The service has an introductory price of $0.54 per hour of audio through the end of the year.

Two models for AI-generated speech

MAI-Voice-2.1 is Microsoft’s new multilingual text-to-speech model. It supports 23 languages and 26 locales. The company says one synthetic voice can speak across these languages while retaining its identity and using a locally appropriate accent.

The standard model costs $22 per million characters. MAI-Voice-2.1-Flash targets applications that need faster responses or handle high volumes. Microsoft says it can generate up to 45 seconds of audio with 150 milliseconds of end-to-end latency. It costs $15 per million characters.

Both voice models offer voice cloning from a short audio sample and include consent-related safeguards, Microsoft says. The company makes the models available through Microsoft Foundry, MAI Playground, Vercel, and other partners. It also presents a demo called Chatter that combines transcription and voice generation in a live AI agent.