Fermion's Phonon 2 writes down an hour of English speech in about 20 seconds on a MacBook Air
At the end of September, a lab called Fermion Research released Phonon 2, an open speech-to-text model for English. It's a 164 megabyte download, and Fermion says it turns an hour of audio into text in about 20 seconds on a MacBook Air.
The one recording Fermion ships with Phonon 2, and what it wrote
He could wait no longer
He could wait no longer.
Fermion publishes exactly one recording: a two-second clip that comes with its code, to check the install works. Here it is. We ran Phonon 2 on it ourselves, and it wrote it down exactly.
Nearly as accurate as the 2.5 GB Nvidia model it learned from, at a fifteenth of the size
Here's Fermion's chart. Higher means fewer words wrong; further left means a smaller download. Phonon 2 gets about 5.2% of words wrong, averaged over seven test sets. The model it was trained from, Nvidia's Parakeet, gets just under 5%, from a download about 15 times bigger. Fermion says every open model that scores better is at least 5.8 times its size.
A meeting recording, stammer included
because I'm I'm just you know I go into the lab and I say right what am I gonna change today
'Cause I'm I'm just, you know, I go into the lab and I say, Right, what am I gonna change today?
Meetings are the hard case: people talk over each other, stop and restart. Fermion tests on AMI, a set of recorded meetings. Here's a clip from it. Phonon 2 kept the stammer, "I'm I'm", the way the speaker actually said it.
And one it got wrong: “disfluencies” came out as “this fluencies”
probably not but we might need them for disfluencies if we try do it in audio
Probably not, but we might need them for this fluencies. If we try to do it in audio.
Not every clip comes out clean. Listen to this one from the same meeting. The speaker said "disfluencies", a word for exactly those stumbles. Phonon 2 heard "this fluencies".
On meetings it edges out its teacher by a twentieth of a point, and a rival does better
Words wrong on AMI, Fermion's Table 1. Phonon 2's and Parakeet Redux's rows are Fermion's own scoring with the Open ASR Leaderboard's code; the teacher's is the leaderboard's published row.
Fermion says Phonon 2 beats its teacher on meetings. Here are the numbers from its own table. Phonon 2 gets about 9.4% of words wrong on AMI. Its teacher also gets about 9.4%, a twentieth of a point worse. And Parakeet Redux, another small model of about the same size, does better than both, at about 9.2%.
174 times faster than real time on a MacBook Air, by Fermion's measure
Fermion's measurements, Tables 2 and 3: an M5 MacBook Air's graphics chip running Apple's MLX; the runtime comparison is 20 dictations, 797 seconds of speech, on the same machine.
Speed is the other half of the pitch. The download is 164 megabytes, against two and a half gigabytes for its teacher. Fermion measured it at 174 times faster than real time on a MacBook Air, so an hour of audio takes about 20 seconds. That's about 1.7 times the next fastest way to run speech recognition it tried on the same laptop. These are Fermion's own measurements.
Phonon 2 is free to download, for English, under CC BY 4.0
Dated 29 September 2026 in the post's metadata (the page shows no date); the weights went up on Hugging Face on 28 September. The recordings are Fermion's LibriSpeech quickstart clip and AMI test audio (CC BY 4.0); Phonon 2's transcripts are from our own run on 6 October.
Phonon 2 came out at the end of September 2026. The weights are under the same Creative Commons licence as the Nvidia model they come from, so you can use them with credit, and one Python package, fermion research, runs it on Macs, Linux, Windows and Nvidia graphics cards. The post and the weights are linked below.
