Details
- Microsoft AI announced MAI-Transcribe-1.5, a new speech-to-text model built for production transcription across messy real-world audio.
- The model supports 43 languages, expanding from 25 in the earlier version, while Microsoft says it maintains best-in-class accuracy.
- Microsoft says MAI-Transcribe-1.5 sits on the Artificial Analysis accuracy-speed frontier and can transcribe up to five times faster than competing models on long audio.
- The model adds keyword biasing, letting users supply domain-specific terms so the system can better recognize names, jargon, and specialized vocabulary.
- Microsoft says it handles noisy environments, multiple speakers, accents, and rapid speech, and is being integrated into Copilot, Teams, GitHub, and Dynamics 365 Contact Centre.
- It is available in Microsoft Foundry and OpenRouter at $0.36 per hour of audio, with official Microsoft documentation also describing preview access through Azure speech APIs.
- Microsoft says the model can transcribe an hour of audio in under 15 seconds and ranks near the top on public transcription benchmarks.
Impact
The launch strengthens Microsoft’s position in speech infrastructure by combining accuracy, latency, and pricing in a single model rather than forcing buyers to trade one for another. The expanded language coverage and keyword biasing make it more attractive for enterprise transcription workflows, especially contact centers, productivity apps, and domain-specific audio. Against rivals such as OpenAI, Google, ElevenLabs, and Alibaba, the move appears aimed less at novelty than at lowering deployment friction and widening Microsoft’s share of production speech workloads.
