Microsoft has launched its first real-time streaming speech-to-text model, MAI-Transcribe-2-Streaming, which supports 60 languages and automatic language detection. The model is designed for live conversations and real-time captioning, allowing applications to process intent or trigger tools before a user finishes speaking.
According to Artificial Analysis tests published on September 28, the model achieved a word error rate of 2.50% and a final transcription latency of 0.13 seconds. These results place it at the top of a leaderboard featuring 28 streaming models, outperforming Grok Voice Transcribe 2.0 Streaming at 2.73% and ElevenLabs Scribe v2 Realtime at 3.59%.
The model generates initial text segments within approximately 100 milliseconds of receiving audio, with full text appearing in as little as 320 milliseconds. Developers can access the service via Microsoft Foundry, MAI Playground, and OpenRouter. Pricing is currently set at $0.54 per hour or $9 per 1,000 minutes.
This release complements the existing non-streaming MAI-Transcribe-2 model, which is available in Azure Speech public preview for $0.10 per hour. While the non-streaming version is optimized for batch processing of recordings and meeting transcripts, the new streaming variant focuses on immediate output for live interactions.
