Microsoft builds own AI models: MAI-Transcribe beats Whisper

Microsoft's new MAI team (under Mustafa Suleyman, ex-DeepMind) has introduced three models:

MAI-Transcribe-1 — Speech to text in 25 languages, faster and more accurate than OpenAI's Whisper. Ideal for meeting minutes.

MAI-Voice-1 — Text to speech, generates 60 seconds of audio in one second. Natural-sounding voices.

MAI-Image-2 — Image generation as an alternative to DALL-E.

What this means: Microsoft is distancing itself from pure OpenAI dependency. For users this means more choice and potentially lower prices on Azure.