What is Phonon-2?
Phonon-2 is an English speech recognition model from Fermion Research, released on Hugging Face on 28 September 2026. It is a compressed version of NVIDIA's Parakeet TDT 0.6B v3: quantisation-aware training stores each encoder weight at one of five learned levels, about 2.1 bits, which cuts the download from 2,508MB to 164MB.
It runs through MLX on Apple Silicon, on the Neural Engine via a Core ML build, on Linux and Windows CPUs, on NVIDIA GPUs through Docker, and on Android arm64 through native libraries. It is the model inside Detta, Fermion's free on-device dictation app for Apple Silicon Macs.
The weights keep Parakeet's CC-BY-4.0 licence, so commercial use is allowed with attribution; Fermion's command line and code are Apache 2.0.
How much does Phonon-2 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | CC-BY-4.0, attribution required |
| Detta (Mac) | Free | Apple Silicon, macOS 26.2 or later |
Weights free under CC-BY-4.0; CLI and code Apache 2.0; Detta app free on Apple Silicon Macs.
Is Phonon-2 worth it?
Phonon-2's achievement is compression, not a new model, and the numbers hold up well. Across the Open ASR Leaderboard's seven English sets it averages 5.21% word error against 4.96% for the full Parakeet it came from, a quarter of a point lost for a 15-fold smaller file. It beats Moondream's similarly sized Parakeet Redux (178MB, 5.69%), and Whisper large-v3-turbo (1,618MB, 6.58%), and on parliamentary speech (VoxPopuli, 2.46 against 3.19) it beats the original.
Those figures are Fermion's, but they were produced with the leaderboard's own code on the full test sets, and the comparison rows marked as leaderboard figures can be checked. The losses concentrate where you would expect from heavy quantisation: earnings calls (6.96 against 5.85) and LibriSpeech other (3.92 against 3.13), the harder, noisier sets.
Speed is the second reason to care. Fermion reports an hour of audio in about 20 seconds on an M5 MacBook Air (174 times real time), 143 times on eight Zen 5 cores and 6,680 times on one H100 in batches of 128. CPU throughput at that level means a laptop or a small cloud instance can transcribe without a GPU.
The limits are those of Parakeet plus one: English only, punctuation and casing conventions inherited from NVIDIA, no speaker labels, and a download size that is not the runtime memory, since engines expand the weights when they load. Fermion is a small company that does not publish where it is based. For multilingual or diarized work, Nemotron 3.5 Transcribe is the better model; for a 164MB English model that ships inside an app, Fermion's claim to the most accurate open model under 900MB fits every comparison row it publishes.
Where it earns its place
An iOS and macOS developer building a note-taking app
Add offline English dictation without a cloud bill
Bundle the Core ML build for the Neural Engine; a 164MB model fits inside a normal app download, and transcription never leaves the device, which simplifies the privacy policy.
A backend engineer at a podcast-hosting company
Transcribe 5,000 hours of English episodes a month without GPUs
At 143 times real time on 16 vCPUs, one instance clears about 143 hours of audio an hour, so the backlog needs roughly 35 instance-hours a month on the CPU Docker image.
An Android developer at a field-service software firm
Let technicians dictate job notes with no signal
Fermion publishes arm64-v8a CPU engine libraries for API 26 and up, with a C header; word timings come back with --json on desktop, and the CC-BY licence only requires attribution.
Before you commit
- Setup
- pip install fermion-research plus mlx packages on Apple Silicon (Python 3.10 or newer), then phonon transcribe recording.wav; or one docker run with the CPU or CUDA image.
- Integrations
- MLX, Core ML (Neural Engine), Linux and Windows CPU engine, CUDA Docker image, Android native libraries, a local server documented by Fermion, and the Detta Mac app.
- Data and privacy
- Runs entirely on your hardware; nothing is sent to Fermion. Weights CC-BY-4.0 like NVIDIA's original, with a NOTICE listing the changes; code Apache 2.0.
- Cost at scale
- No GPU needed: about 143 times real time on eight Zen 5 cores. On one H100 in batches of 128, Fermion reports 6,680 times real time, about 111 hours of audio a minute.
- Limits
- English only; output conventions are Parakeet's; no diarization; runtime memory is larger than the 164MB download because weights are expanded at load.
What’s good
- 5.21% average English WER from a 164MB download, against 4.96% for the 2.5GB original
- 174 times real time on an M5 MacBook Air; 143 times on eight CPU cores
- Runs on MLX, Core ML, Linux and Windows CPUs, CUDA and Android
- CC-BY-4.0 allows commercial use with attribution
What’s not
- English only
- Loses about a point of WER on earnings calls versus the original
- No speaker labels or translation
- Accuracy and speed figures are Fermion's own
Best for
App developers who need accurate English speech-to-text on a laptop, phone or CPU server without shipping a multi-gigabyte model.
Phonon-2 alternatives
Phonon-2: common questions
- Is Phonon-2 worth it?
- Phonon-2 is Parakeet at a fifteenth of the size and nearly the same accuracy, which makes good English transcription practical inside an app download. We rate it 4.1 out of 5.
- How much does Phonon-2 cost?
- Weights free under CC-BY-4.0; CLI and code Apache 2.0; Detta app free on Apple Silicon Macs.
- What are the downsides of Phonon-2?
- English only; loses about a point of WER on earnings calls versus the original; no speaker labels or translation; accuracy and speed figures are Fermion's own.
- Who is Phonon-2 best for?
- App developers who need accurate English speech-to-text on a laptop, phone or CPU server without shipping a multi-gigabyte model.
- How much does Phonon-2 cost for a team or in production?
- No GPU needed: about 143 times real time on eight Zen 5 cores. On one H100 in batches of 128, Fermion reports 6,680 times real time, about 111 hours of audio a minute.
- Is Phonon-2 safe for company data?
- Runs entirely on your hardware; nothing is sent to Fermion. Weights CC-BY-4.0 like NVIDIA's original, with a NOTICE listing the changes; code Apache 2.0.
- What are the best Phonon-2 alternatives?
- The closest alternatives we have reviewed are Parakeet TDT 0.6B v3, Whisper large-v3-turbo and NVIDIA Nemotron 3.5 Transcribe.
Answers checked Oct 31, 2026.
Make Phonon-2? Show this rating on your site
