Microsoft has launched MAI-Transcribe-2-Streaming, a real-time speech-to-text model capable of processing 60 languages with continuous detection. The model provides partial transcriptions within 100ms of receiving audio, allowing AI agents to process data mid-sentence. Internal testing by Microsoft indicates that the model generates text twice as fast as its closest competitor.
The model is evaluated on the AA-WER Streaming index, which utilizes 8 hours of audio data from sources including AA-AgentTalk, VoxPopuli, and Earnings22. Latency is measured using SileroVAD from the end of speech. Microsoft positions the model on the accuracy versus latency Pareto frontier, noting that initial partial transcriptions maintain high accuracy.
Pricing for the service is set at $0.54 per hour of audio, which is an introductory rate valid until the end of 2026. This equates to $9.00 per 1,000 minutes. For comparison, the batch version of the model costs $0.10 per hour. The streaming model is priced higher than offerings from xAI and Meta, while matching estimated rates from Google.
Developers can integrate the model via the Realtime API, compatible with OpenAI WebSockets, or through the Azure Speech SDK. The model is accessible via the MAI Playground, Vercel, and Azure Voice Live, with LiveKit support expected soon. It is designed to pair with MAI-Voice-2.1-Flash for full voice interaction loops.
