Microsoft launches speech-transcribing AI for 10 cents an hour — and it works 10 times faster than its OpenAI counterpart

Microsoft AI department releases artificial intelligence model mai-transcription-2designed specifically for speech recognition. The developers claim that it works faster, more accurately, and is cheaper than similar solutions from OpenAI, Google, or ElevenLabs. The fee for transcribing one hour of recording is 10 cents.

Image source: microsoft.ai

The new Microsoft MAI-Transcribe-2 supports speech transcription in 60 languages, including Russian – by comparison, June’s MAI-Transcribe-1.5 supports 43 languages, and the first version released in April only supported 25 languages. It runs on the Microsoft Foundry platform for developers and the MAI Playground for testing. The company optimized the model to handle low-quality audio, which is what is needed in practice – with background noise, low recording quality and overlapping cues. The system tags speakers and places timestamps at the word level, allowing for searching, as well as easier editing and synchronization with videos. There is also the ability to enter keywords: you can enter a list of medicines, product codes or proprietary names so that the model does not distort them while working. There is verbatim output with filler words and cleaned recordings; speech transcription supports fast switching between languages, even mid-phrase.

In the FLEURS ratings for 60 languages, Microsoft’s MAI-Transcribe-2 model ranked first with an average error rate of 5.2% for single-word recognition. June’s MAI-Transcribe-1.5 was 3.7%, but this dynamic doesn’t mean the model is getting worse – probably because coverage expanded from 43 to 60 languages. In the independent human analysis ratings, compared with June, the new model rose from third to second place, beating ElevenLabs Scribe v2 and second only to Alibaba Fun-Realtime-ASR-preview. Finally, Microsoft claims that the new MAI-Transcribe-2 is ten times faster than OpenAI GPT-Transcribe, seven times faster than ElevenLabs Scribe v2, and five times faster than Google Gemini 3.5 Transcribe. If the first version of the speech transcription model released by Microsoft in April that supports 25 languages ​​costs $0.36 per hour for recording, then the new version that supports 60 languages ​​costs $0.10 for the same amount of data – however, Microsoft said that this is the starting price, and it is unknown whether it will change in the future.