Microsoft AI releases MAI-Transcribe-2-Streaming for real-time transcription and MA

Summary

Microsoft AI has introduced MAI-Transcribe-2-Streaming, a real-time transcription model supporting 60 languages with an initial partial latency of approximately 100 ms. This is accompanied by MAI-Voice-2.1, a multilingual text-to-speech model capable of cloning voices from just a few seconds of audio. Both models are available at introductory prices via Microsoft Foundry, MAI Playground, and OpenRouter.

Why it matters

Transcription latency under 100 ms falls below the threshold that allows voice agents to interrupt mid-sentence and respond, marking a significant step toward AI with natural conversational abilities. The 23-language single-voice text-to-speech with local accents and the 150 ms latency Flash version address a market where multilingual consistency has long been a gap. The $0.54 per hour transcription price is lower than many real-time STT offerings currently available during the introductory period.

Highlights

• 60-language real-time STT with initial partial results in ~100 ms; Microsoft claims #1 accuracy on Artificial Analysis
• MAI-Voice-2.1-Flash: 150 ms TTS latency, 150 ms Flash version, single and consistent voice in 23 languages, $15/1M characters (down from $22)
• Voice cloning from a few seconds of reference audio with vague 'built-in safety measures'; approximately 50% of 4,000 test listeners perceived the output as human

Critical take

All headline metrics, including the top accuracy ranking, 100 ms latency, and the 50% human illusion test, come from Microsoft's own presentation. The article provides no independent benchmarks or methodology to verify these claims. The 'built-in safety measures' regarding voice cloning are described in a single, ambiguous sentence, leaving consent and misuse prevention mechanisms entirely unclear.