Microsoft ships MAI-Transcribe-2

Microsoft AI shipped MAI-Transcribe-2 on September 3, a speech-recognition model the company says takes the top spot on the 60-language FLEURS benchmark at a 5.2% average word error rate. Microsoft’s own comparisons put inference speed at roughly 10 times OpenAI’s GPT-Transcribe, seven times ElevenLabs’ Scribe v2, and five times Google’s Gemini 3.5 Transcribe, alongside new speaker diarization and word-level timestamps the prior generation didn’t have.

Pricing is the headline: $0.10 per audio hour through December 31, in public preview via Azure Speech. Microsoft’s first model in this line, shipped five months ago, charged $0.36 an hour, so this is roughly a 72% cut from its own predecessor, not just from competitors. It’s the latest move in a transcription market where vendors are increasingly competing on price and language coverage rather than accuracy alone, following Cohere’s open-sourced Arabic speech model earlier this year.

For builders running high-volume transcription (call centers, meeting tools, captioning pipelines) the combination of speed and price is worth testing against whatever you’re using now. The catch is the word “introductory”: Microsoft hasn’t committed to $0.10/hour past year-end, and a promotional rate that undercuts the market is also a rate that can reset once it has pulled in switching customers. Budget accordingly rather than locking in a workflow around a price that’s guaranteed for four months.