Mai Transcribe 1 Everything You Need To Know

Microsoft's MAI-Transcribe-1 has a 3.0% word error rate on AA-WER, the Artificial Analysis speech to text accuracy benchmark, placing it fourth (artificialanalysis.ai). Word error rate is roughly the share of words a model gets wrong against a reference transcript, so lower is better. The three models ahead are ElevenLabs' Scribe v2 at 2.3%, and Mistral's Voxtral Small and Google's Gemini 3.1 Pro High, both at 2.9%, which leaves MAI-Transcribe-1 0.1 point behind second place.

Its distinguishing number is speed. It processes about 69 seconds of audio for every second of compute, which Artificial Analysis says makes it the fastest of the top five models by accuracy. The model comes from Microsoft AI's Superintelligence team, supports 25 languages including English, French, Arabic, Japanese and Chinese, and is in public preview through Azure Speech on Microsoft Foundry at $6 per 1,000 minutes of audio. Two months later Microsoft followed it with MAI-Transcribe-1.5, which Artificial Analysis measured at 2.4% and about 276 times real time.