Microsoft releases three MAI models
On 2 April 2026 Microsoft AI released three of its own models in Microsoft Foundry at once: MAI-Transcribe-1 (speech-to-text), MAI-Voice-1 (text-to-speech) and MAI-Image-2 (image generation), replacing third-party models at both ends of its voice stack with its own.
Why it matters
A company whose products had relied on Whisper and other third-party speech models released, for the first time by this date, its own models for both speech recognition and speech synthesis at once, alongside a new generation of its own image generator. An editorial assessment.
MAI-Transcribe-1: by the FLEURS benchmark, a mean word error rate of 3.9% across the top 25 languages - below GPT-Transcribe (4.2%), Scribe v2 (4.3%), Gemini 3.1 Flash (4.9%) and Whisper-large-v3 (7.6%); by the company's own statement, batch transcription speed is 2.5x the existing Azure Fast offering, and among the top 25 languages the model ranks first by FLEURS in 11. MAI-Voice-1: generates 60 seconds of audio in one second; the same day, Microsoft Foundry added the ability to create a custom voice from a few seconds of audio. MAI-Image-2: at least twice as fast at image generation on Foundry and Copilot at similar quality compared with the previous generation, based on real production-traffic data; WPP is named as an early major client. All three models are available immediately in Microsoft Foundry and the MAI Playground, priced from $0.36/hour (MAI-Transcribe-1), $22 per million characters (MAI-Voice-1) and $5/$33 per million tokens of text input/image output (MAI-Image-2). What the record does not claim. The comparative WER and speed figures are the company's own claims about its own tests, without independent replication. The announcement page has been edited since April (marked 'Updated as of June 9, 2026'); by an archived copy from 3 April 2026, the set of models and the figures used here were already present in the announcement at that time - the June edit did not add or change them.