Microsoft AI has introduced MAI-Transcribe-2-Streaming, a real-time speech transcription model supporting 60 languages. According to the company, the model achieved top accuracy rankings on Artificial Analysis and returns partial transcription outputs in just over 100 milliseconds, allowing voice agents to process input while users are speaking. The model is currently offered at an introductory rate of $0.54 per hour of audio.

In addition to transcription, Microsoft released two text-to-speech models: MAI-Voice-2.1 and MAI-Voice-2.1-Flash. MAI-Voice-2.1 supports native accents across 23 languages using a single voice clone, while the Flash variant achieves 150-millisecond latency at a price of $15 per million characters. Both models can clone voices from a few seconds of audio and include built-in safety controls.

The models are accessible via Microsoft Foundry, MAI Playground, and OpenRouter. Microsoft reported that in human evaluations involving 4,000 participants, approximately half of respondents believed the synthetic voices belonged to a real person.

Why it matters

  • Sub-150ms transcription and voice synthesis latencies enable true real-time, conversational voice agents for customer and enterprise workflows.

  • Low-cost voice cloning models across 23 languages lower the barrier for developers building global voice-first applications.

Source: the-decoder.com