Research
Introducing Phonon-2.5
Open speech recognition in 25 languages from a 156 MB download, and a new English model of the same size. Both run on a laptop CPU at the speed of Phonon-2.
Speech recognition · English · Available nowPhonon-2.5 English
Our most accurate English speech recognition model, in a 158 MB download. 4.82 % average word error on the seven Open ASR Leaderboard sets, better than Phonon-2 on all seven.
- 158 MB
- Download, 178 MB on disk
- 4.82 %
- Seven-set word error rate
- 7 of 7
- Sets better than Phonon-2
- 1.64 %
- LibriSpeech test-clean
- 3.56 %
- LibriSpeech test-other
Overview
Phonon-2.5 English is an open speech recognition model for English in a 158 MB download. Across the Open ASR Leaderboard’s seven English sets it averages 4.82 % word error, against 4.94 % for Underdog Bark 3.0, the other low-bit model of this size, and 5.21 % for Phonon-2. It is better than Phonon-2 on all seven sets, and it is our most accurate English model.
Some workloads want English text and nothing else: dictation, captions, compliance and search pipelines built around English. Phonon-2.5 English is for those. The same file runs on Apple silicon and on Linux (x86-64 and Arm) and Windows CPUs, with no GPU needed, on NVIDIA GPUs and in the CPU and CUDA Docker images. It writes punctuated, capitalized text with a start and end time for every word, and the weights are released under CC-BY-4.0. For 25 languages, Phonon-2.5 is the same size.
Models
Phonon-2.5 English is a single file. Phonon-2.5, the 25-language model, and Phonon-2 remain available beside it.
- Phonon-2.5 EnglishEnglish
- 158 MB download · 178 MB on disk
- Phonon-2.525 languages
- 156 MB download · 178 MB on disk
- Phonon-2English; the earlier model
- 164 MB download · 177 MB on disk
- Language
- English
- Audio inputmicrophone or file
- 16 kHz
- Outputwith a start and end time for every word
- Punctuated, capitalized text
Evaluation
Phonon-2.5 English averages 4.82 % word error on the seven public test sets of the Open ASR Leaderboard, scored with the board’s own code on the full test sets, with Earnings-22 as the leaderboard’s chunked set. Table 1 sets it beside Phonon-2.5, Phonon-2, Underdog Bark 3.0, Parakeet Redux and the 2,508 MB Parakeet TDT 0.6B v3 the Phonon-2 family derives from.
| Model | Download | LS clean | LS other | AMI | Earnings-22 | GigaSpeech | SPGISpeech | VoxPopuli | Average |
|---|---|---|---|---|---|---|---|---|---|
| Phonon-2.5 English | 158 MB | 1.64 | 3.56 | 8.76 | 6.22 | 7.99 | 3.31 | 2.28 | 4.82 |
| Phonon-2.525 languages | 156 MB | 1.60 | 3.41 | 9.12 | 6.72 | 8.04 | 3.31 | 3.23 | 5.06 |
| Phonon-2 | 164 MB | 1.72 | 3.92 | 9.37 | 6.96 | 8.35 | 3.70 | 2.46 | 5.21 |
| Underdog Bark 3.0 | 163 MB | 1.89 | 4.08 | 7.82 | 7.05 | 7.95 | 3.72 | 2.05 | 4.94 |
| Parakeet TDT 0.6B v3† | 2,508 MB | 1.52 | 3.13 | 9.42 | 5.85 | 7.99 | 3.63 | 3.19 | 4.96 |
| Parakeet Redux | 178 MB | 1.94 | 4.35 | 9.16 | 7.90 | 8.62 | 4.01 | 3.87 | 5.69 |
Phonon-2.5 English158 MB
- LS clean
- 1.64
- LS other
- 3.56
- AMI
- 8.76
- Earnings-22
- 6.22
- GigaSpeech
- 7.99
- SPGISpeech
- 3.31
- VoxPopuli
- 2.28
- Average
- 4.82
Phonon-2.525 languages156 MB
- LS clean
- 1.60
- LS other
- 3.41
- AMI
- 9.12
- Earnings-22
- 6.72
- GigaSpeech
- 8.04
- SPGISpeech
- 3.31
- VoxPopuli
- 3.23
- Average
- 5.06
Phonon-2164 MB
- LS clean
- 1.72
- LS other
- 3.92
- AMI
- 9.37
- Earnings-22
- 6.96
- GigaSpeech
- 8.35
- SPGISpeech
- 3.70
- VoxPopuli
- 2.46
- Average
- 5.21
Underdog Bark 3.0163 MB
- LS clean
- 1.89
- LS other
- 4.08
- AMI
- 7.82
- Earnings-22
- 7.05
- GigaSpeech
- 7.95
- SPGISpeech
- 3.72
- VoxPopuli
- 2.05
- Average
- 4.94
Parakeet TDT 0.6B v3†2,508 MB
- LS clean
- 1.52
- LS other
- 3.13
- AMI
- 9.42
- Earnings-22
- 5.85
- GigaSpeech
- 7.99
- SPGISpeech
- 3.63
- VoxPopuli
- 3.19
- Average
- 4.96
Parakeet Redux178 MB
- LS clean
- 1.94
- LS other
- 4.35
- AMI
- 9.16
- Earnings-22
- 7.90
- GigaSpeech
- 8.62
- SPGISpeech
- 4.01
- VoxPopuli
- 3.87
- Average
- 5.69
Speed
Phonon-2.5 English has the architecture, the file format and the engines of Phonon-2, so it runs at about the same speed as Phonon-2, and it will get a lot faster in the coming weeks. For reference, these are Phonon-2’s published figures, from its release post: 174 times realtime on a base M5 MacBook Air, which is an hour of audio in about 20 seconds; 143 times realtime on eight AMD Zen 5 cores; and 6,680 times realtime batched on an H100.
Run it
Python 3.10 or newer. The fermion command line installs from PyPI; name the model as phonon-2.5-en.
$ pip install fermion-research$ fermion transcribe phonon-2.5-en meeting.wav$ fermion transcribe phonon-2.5-en meeting.wav --json$ fermion serve phonon-2.5-en --port 8010
It transcribes files, adds a start and end time for every word under --json, and serves an OpenAI-compatible /v1/audio/transcriptions endpoint.
$ pip install "fermion-research[mlx]"
On Apple silicon it runs the model on the GPU through MLX.
$ pip install --no-deps torch --index-url https://download.pytorch.org/whl/cpu$ pip install fermion-research
On Linux (x86-64 and Arm) and Windows the same package runs the CPU engine; on Linux, install torch from its CPU wheel index first to skip the GPU build.
Weights on Hugging FaceDocumentationfermion-research on PyPIEngines on GitHub
Availability
Open weightsCC-BY-4.0EnglishApple siliconNVIDIA GPUsCPUs
License
The weights are released under the Creative Commons Attribution 4.0 licence, which Phonon-2.5 English inherits from NVIDIA’s Parakeet TDT 0.6B v3; the repository’s NOTICE carries the attribution and lists the changes. The licence permits commercial use, modification and redistribution with attribution. The command line is released under Apache 2.0.
Citation
@misc{fermionresearch2026phonon25english,
title = {Phonon-2.5 English},
author = {{Fermion Research}},
year = {2026},
url = {https://fermionresearch.com/models/phonon-2-5-english/}
}