A new real-time speech-to-text system from Microsoft AI has claimed the top position for precision on the AA-WER Streaming benchmark. The MAI-Transcribe-2-Streaming model achieves a 2.5% word error rate (WER) on final transcripts, ranking first among 38 tested models. It returns results 0.13 seconds after speech ends, displacing the previous leader, SpaceXAI's Grok Voice Transcribe 2.0, which had a 2.7% WER and a 0.49 second latency.

The tool costs $0.54 per hour of audio, or $9 per 1,000 minutes, placing it level with Gemini 3.5 Transcribe Live and above rivals ElevenLabs and Deepgram Flux at $6.50. Non-streaming transcription is available for $0.10 per hour. Mustafa Suleyman, CEO of Microsoft AI, described the model as the most accurate of its kind in the world and urged developers to build agents on the platform.