Microsoft's New Speech Model Undercuts Everyone at 10 Cents an Hour

Microsoft AI has released MAI-Transcribe-2, a speech-recognition model it says is the fastest, most accurate and cheapest on the market, priced at 10 cents per hour of audio through the end of the year. The lab announced the model on September 3, with a public preview running on Azure Speech.

The headline claim is a clean sweep of the usual comparisons. Microsoft says the model ranks first on FLEURS, a Google benchmark that measures transcription accuracy across dozens of languages, with an average word error rate of 5.2 percent over 60 languages. On speed, it puts MAI-Transcribe-2 at up to ten times faster than OpenAI's GPT-Transcribe, seven times faster than ElevenLabs' Scribe v2, and five times faster than Google's Gemini 3.5 Transcribe.

The price is the story

Benchmarks come and go, and every lab picks the ones that flatter it. The number that will get attention is the price. Ten cents an hour is aggressive. Microsoft's first model in this line, shipped only five months ago, charged 36 cents, so the introductory rate is a cut of roughly 72 percent, and it undercuts the transcription APIs from OpenAI, Google and ElevenLabs at the same time.

That introductory price runs through December 31, which is the part worth reading twice. Cheap launch pricing that resets on New Year's Day is now a familiar move, one Google used with its recent Gemini Flash pricing. It buys market share and benchmark headlines up front, and quietly leaves the 2027 rate to be settled later.

What it does

Beyond raw transcription, MAI-Transcribe-2 adds speaker diarization, which labels who said what, along with word-level timestamps, keyword biasing and automatic language identification. It also handles code-switching, the everyday habit of moving between languages inside a single conversation that has long tripped up automatic transcription.

Microsoft made a point of noting that a ten-person team built the model, a detail clearly aimed at the idea that only the largest labs with the biggest budgets can compete at the frontier. Speech recognition is a narrower problem than general reasoning, so a small team landing at the top of a benchmark is less surprising there than it would be for a flagship chatbot. Still, at 10 cents an hour, the pressure lands squarely on rivals who have been charging several times as much for the same job.