Microsoft launches MAI-Transcribe-2 preview at 10 cents per audio hour

Microsoft has put its MAI-Transcribe-2 speech-to-text model into public preview at an introductory price of $0.10 per audio hour through December 31, 2026. That single metered rate covers transcription, speaker diarization and word-level timestamps.

The model runs in Microsoft Foundry and through Azure Speech, and it supports 60 languages, according to Microsoft. The company has not disclosed what the model will cost once the introductory period ends, or when it will reach general availability.

Diarization and word-level timestamps

Speaker diarization splits a recording by voice, so a transcript marks which speaker delivered each passage. Word-level timestamps attach start and end times to individual words, which supports work such as caption alignment, clip selection and search inside long recordings.

The release also extends Microsoft’s lineup of task-specific MAI models, which already includes the MAI-Cyber-1-Flash cybersecurity model.

Microsoft’s benchmark and price claims

Microsoft says the model ranks first on FLEURS, a multilingual speech-recognition benchmark, and second on the Artificial Analysis word-error-rate leaderboard. It also places MAI-Transcribe-2 first on that leaderboard’s accuracy-latency Pareto frontier, meaning no listed competitor is both more accurate and faster.

Microsoft describes the model as the lowest-priced transcription model on the market. The assigned evidence does not independently verify the benchmark placements or that market-wide pricing claim.

The announcement does not give per-language accuracy results, a diarization error rate, regional availability or preview capacity limits, and it does not establish whether MAI-Transcribe-2 supports real-time streaming transcription.