Microsoft announces its first streaming transcription model

Summary

Microsoft AI has introduced its first streaming transcription model, MAI-Transcribe-2-Streaming, along with two text-to-speech (TTS) models, MAI-Voice-2.1 and its Flash variant, as a public preview on Microsoft Foundry. These models are designed with a focus on developing real-time voice agents.

Why it matters

This move places Microsoft in direct competition with OpenAI's GPT-Realtime-Whisper, Google's Gemini 3.8 Live, and Meta's open-source Omnilingual ASR in the rapidly evolving real-time voice AI space. A provisional transcription latency of under 100 milliseconds and an aggressive starting price ($0.54 per hour) indicate Microsoft's intent to lock in enterprise voice agent workloads to the Foundry platform. For developers, the bundling of ASR, TTS, and voice cloning components reduces integration barriers for multilingual speech products.

Highlights

• Approximately 100 milliseconds of provisional transcription latency with automatic detection across 60 languages; claims to rank #1 in the Artificial Analysis streaming ASR benchmark
• MAI-Voice-2.1-Flash generates 45 seconds of audio with 150 milliseconds of latency. It is approximately 60% cheaper than similar TTS models ($15 per million characters)
• Voice cloning is limited by an approval and review process. Although the service is available in the Japan East region, Japanese is not included in the default voice list

Critical take

The highlighted 'Chatter' demo accepts Japanese input but can only respond in English. This exposes a practical gap in the 60-language claim within the Japanese market, where the model is regionally hosted. Additionally, the #1 accuracy rate is self-cited from Artificial Analysis. The $0.54 per hour ASR price is an explicit introductory rate valid only through the end of the year, leaving long-term cost competitiveness uncertain.