Microsoft Ships First Streaming Transcription Model, Completing Voice Agent Pipeline on Foundry

On October 1, 2026, Microsoft launched MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash through Foundry and Azure AI Speech. The streaming model claims 0.13s end-of-speech latency versus 0.49s for Grok Voice Transcribe 2.0 and costs $0.54 per audio hour against $0.10 for the batch model.

• End-of-speech latency of 0.13s versus 0.49s for Grok Voice Transcribe 2.0
• Priced at $0.54 per audio hour through December 31, 2026, versus $0.10 batch
• Supports 60 languages with auto-detection, claims 2.5% WER
• All three models in public preview with no SLA and no streaming diarization