Microsoft launched MAI-Transcribe-2-Streaming for real-time transcription. It ranks first for accuracy on Artificial Analysis. It returns partial transcripts in over 100ms. It supports 60 languages with automatic detection. It costs $0.54 per hour through year-end. Microsoft also released MAI-Voice-2.1 for multilingual text-to-speech. It supports 23 languages and 26 locales with a single voice. It is priced at $22 per 1M characters. MAI-Voice-2.1-Flash offers 150ms latency and costs $15 per 1M characters. It is 55% faster and 60% cheaper than comparable models. Both voice models support cloning with consent guardrails. They enable voice agents to act mid-sentence and maintain conversational speed.
What changed
Previous transcription models returned text only after full audio input.
What it unlocks
Real-time voice agent interactions with mid-sentence reasoning and tool use.
- $0.54 per hour of audio for MAI-Transcribe-2-Streaming
- $22 per 1M characters for MAI-Voice-2.1
- $15 per 1M characters for MAI-Voice-2.1-Flash
- 150ms end-to-end latency for MAI-Voice-2.1-Flash
What you need to act on it
- access via Microsoft Foundry, OpenRouter, MAI Playground, Vercel, or LiveKit
Sources