Phonon-1 is an open speech recognition model for English. It downloads in 415 MB, runs on a laptop or a datacenter GPU, and transcribes an hour of audio in about two and a half minutes. It was trained at 2.4 bits per weight from the start, and it is the second model in the lab’s low-bit lane after Neutrino-1.
Across five real-world benchmarks (meetings, earnings calls, web video, financial speech, talks), no downloadable model we could find is both smaller and more accurate. Against the model it was trained from, Phonon-1 keeps 96.7 to 99.6 % of its teacher’s word accuracy in 29 % of the bytes, about a fifth more relative word errors. The output arrives punctuated and capitalized from the model itself, and the weights are open under Apache 2.0, released today alongside a smaller build, Phonon-1 Micro.
Benchmarks
| Benchmark | Phonon-1415 MB | Phonon-1 Micro285 MB | Parakeet-0.6B 4-bit637 MB | Moonshine base248 MB | Whisper large-v3-turbo1,619 MB | Whisper small967 MB | wav2vec2-large1,262 MB | Qwen3-ASR teacher1,569 MB |
|---|---|---|---|---|---|---|---|---|
| LibriSpeech test-clean | 2.640 | 3.002 | 2.186 | 3.417 | 2.10 | 3.4† | 2.8† | 2.235 |
| LibriSpeech test-other | 5.699 | 6.511 | 3.937 | 8.262 | 4.07 | 7.6† | 6.3† | 4.618 |
| TED-LIUM | 3.421 | 3.878 | 2.829 | 5.272 | – | – | – | 2.889 |
| SPGISpeech | 4.163 | 4.858 | 4.104 | 5.731 | 2.79† | – | 13.31† | 3.074 |
| VoxPopuli | 8.394 | 9.177 | 6.345 | 10.470 | 11.22† | – | – | 7.151 |
| GigaSpeech | 11.396 | 11.882 | 9.614 | 12.114 | 8.52† | – | – | 9.321 |
| Earnings-22 | 12.571 | 14.771 | 11.190 | 17.872 | 11.07† | – | 36.28† | 11.188 |
| AMI | 13.084 | 14.094 | 12.723 | 17.790 | 15.16† | – | – | 12.560 |
| Macro, eight benchmarks | 7.67 | 8.52 | 6.62 | 10.1 | – | – | – | 6.63 |
Speed
File transcription runs at a median 23.9× realtime across nine corpora on a base M5 MacBook Air. In live dictation, Phonon-1 completes the median utterance in 262 ms and 95 of every 100 utterances within 629 ms; the fastest build of the family measures 203 ms at the median and 446 ms at the 95th percentile.
The streaming figures describe the mode a user actually runs: streamed decode is verified byte-identical to batch decode on 400 of 400 utterances. The same weights run end to end on NVIDIA GPUs at the same measured accuracy, with the runtime and Docker image in the GitHub repository and long-form and streaming serving rolling out in the container.
Phonon-1 Micro
Phonon-1 Micro is the smallest build of the family: a 285 MB download, 331 MB on disk, state of the art at its size. It measures 3.002 % on LibriSpeech test-clean and 6.511 % on test-other, and it is more accurate than Moonshine base, the nearest smaller model, on all eight benchmarks in the table, a macro of 8.52 against 10.1. The weights are at FermionResearch/Phonon-1-Micro, with specifications on the model page.
Run it
The command line installs from PyPI, pulls the model on first run, and can serve a local transcription endpoint. The same weights run on a Mac (via MLX) or an NVIDIA GPU.
$ pip install fermion-research$ fermion transcribe recording.wav$ fermion serve
$ docker run --rm --gpus all ghcr.io/fermionresearch/phonon-cuda:latest \transcribe recording.wav
The container transcribes recordings of any length, streams over WebSocket, and serves the same endpoint as fermion serve.
Availability
Phonon-1 is available now under Apache 2.0: 415 MB to download, 455 MB on disk, with Phonon-1 Micro at 285 and 331 MB. Both models transcribe English from 16 kHz audio. The base model, Qwen3-ASR-0.6B, is Apache 2.0; the training corpora are CC-BY-4.0 and CC0, and the LibriSpeech and MLS attribution notice travels with any redistribution.
A dictation application for macOS built on Phonon is coming soon.
Weights and specifications: FermionResearch/Phonon-1 on Hugging Face, with the model page here. Command line: pip install fermion-research.
