Microsoft launches its first real-time streaming speech transcription model, covering 60 languages

Microsoft has released its first real-time streaming speech transcription model, MAI-Transcribe-2-Streaming, which outputs text continuously as speech happens. It covers 60 languages and supports automatic language detection. The model generates its first partial text within about 100 milliseconds of receiving audio, then keeps correcting it and quickly delivers stable results after a sentence ends.

According to Microsoft's real-time evaluation, text can appear as fast as about 320 milliseconds after speech begins. In Artificial Analysis testing, the model achieved a final transcription latency of 0.13 seconds and a final word error rate of 2.50%, ranking first among 28 streaming speech-to-text models. It beat Grok Voice Transcribe 2.0 Streaming at 2.73% and ElevenLabs Scribe v2 Realtime at 3.59%.

The model is currently available at a promotional price of $0.54 per hour. Developers can access it through Microsoft Foundry, MAI Playground, OpenRouter and other channels. The streaming mode suits real-time conversations and captions, allowing voice apps to understand intent, start reasoning or call tools before a user finishes speaking.