Microsoft has released a speech recognition model priced at around $0.10 per hour of audio until the end of 2026, down from around $0.36 for the previous version five months ago — a drop of roughly 72%. Coverage expands from 43 to 60 languages, with speaker identification and word-level timestamps. The company claims an average error rate of 5.2% on a benchmark multilingual test.
We have covered voice as the primary interface and native voice models. Here is the most mundane and perhaps most useful piece of the puzzle.
What ten cents an hour changes
Let's do the maths, because it is telling. Transcribing a one-hour meeting costs ten cents. A day of interviews, less than a euro. Ten years of audio archives from a local radio station, a few dozen euros.
At that level, transcription stops being a trade-off. You no longer ask whether it's worth the cost: you transcribe everything, systematically, because not doing so buys you nothing.
This is the kind of shift we have observed elsewhere. When an operation becomes cheap enough, it isn't its volume that grows — it's its nature that changes. You move from a service you order to a function running in the background.
The uses this unlocks
Accessibility. This is the most indisputable benefit, and we stress it every time. Systematic captioning of all audio content becomes realistic for any organisation, including non-profits. We explored this in our article on accessibility.
Archives. Decades of recordings — audiovisual, legal, scientific, family — become searchable by text query. That is a major shift in access for collections that sit dormant for lack of means to index them.
Less-served languages. The jump from 43 to 60 languages matters more than the price for parts of the world. We wrote about the languages major models ignore that the imbalance lies in technical details. Language coverage is one of them.
When transcription costs nothing, everything gets transcribed. Including what you'd rather leave unspoken. A meeting, a hallway conversation, a phone call becomes a searchable, storable, exploitable document. We wrote this about voice as an interface: installing permanent microphones in our spaces is a collective decision no one has made. A transcription cost near zero removes the last practical barrier to systematic recording.
The caveat on the figures
Two usual precautions apply.
The price is promotional. It is announced until the end of 2026. We saw this summer what a promotional price becomes when it expires. Don't build an economy on it without knowing what you'll do next.
The average error rate hides huge variations. A figure aggregated across sixty languages says nothing about performance on yours, nor on a noisy recording, a regional accent, or several people speaking at once. As always, the only measure that matters is the one you run on your own files.
What to take away
This announcement is nothing spectacular, and it is representative of what is actually happening. While we debate AGI, a mundane technical component has become so cheap that it changes the practices of an entire profession.
This is exactly what we noted in our article on progress: real benefit rarely sits where the measuring instruments are looking. A model that solves open mathematical problems makes headlines. A ten-cent-an-hour transcription will let thousands of deaf people access content that was closed to them. Both matter, and one of the two will matter sooner.