Breaking

Microsoft Launches MAI-Transcribe-2-Streaming and MAI-Voice-2.1 Models for Real-Time Voice Agents

October 1, 2026 at 14:16 EDT

Foundation Models
AI Agents
Media Generation

Microsoft AI launched MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash on October 1, expanding its in-house audio stack for low-latency conversational agents. The new streaming transcription model is positioned for real-time speech understanding, while the two text-to-speech models add multilingual voice generation options for higher-fidelity and faster, lower-cost use cases.

October 1, 2026 · Microsoft AI

A faster audio stack for voice agents

Microsoft launched MAI-Transcribe-2-Streaming with two MAI-Voice-2.1 models, combining real-time listening with multilingual speech generation for conversational AI.

transcription languages

voice languages

100+ ms

first partial transcripts

$0.54

per audio hour through 2026

Voice generation price comparison

$22

MAI-Voice-2.1 per 1M chars

$15

MAI-Voice-2.1-Flash per 1M chars

Flash targets lower-latency, high-volume speech at a lower listed character price.

What improves

Real-time partial transcripts, multilingual voices and faster output give agents more time to reason, call tools and respond naturally.

What teams must test

Accuracy in noisy audio, domain terminology, the quality-speed tradeoff and consent requirements for voice cloning remain practical checks.

The agent loop

Hear→Understand→Decide→Speak

Microsoft is trying to shorten both the listening and speaking sides of the conversation window.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.