Two ways to keep audio inside
- On the device: Detta, our dictation app for the Mac, recognises speech on the Mac's Neural Engine. With Wi-Fi off, dictation keeps working. Detta asks for a Google sign-in, and sharing dictations to improve our models is on by default with an off switch in Settings; with sharing off, no dictation is uploaded.
- On your servers: fermion serve runs Phonon-2.5 (25 languages, 156 MB) or Phonon-2.5 English on CPU or GPU machines in your cloud or data centre, with an OpenAI-compatible endpoint and an API key.
Accuracy you can check
The models are open and scored on public test sets: Phonon-2.5 English averages 4.82 % word error on the seven Open ASR Leaderboard sets; Phonon-2.5 averages 11.71 % across 25 languages on FLEURS. Your team can re-run the numbers on your own recordings before deploying.
Capacity on ordinary hardware
Published Phonon-2 figures (Phonon-2.5 shares its size and architecture): 143 times realtime on eight AMD Zen 5 cores, 174 times realtime on an M5 MacBook Air GPU, 6,680 times realtime batched on an H100. A single eight-core CPU server transcribes more than a hundred hours of audio per hour.
Custom vocabulary
Clinical, legal and financial speech is full of terms a general model has not seen. The fermion runtime takes a hotword list per call (names, drugs, products) that biases recognition toward the listed terms without changing the transcript when they are not said; Detta has a dictionary for names and terms.
fermion transcribe phonon-2 visit.wav --hotwords @drug-names.txtQuestions
- Can speech recognition run without any cloud service?
- Yes. Phonon runs on your own CPU or GPU servers through the fermion runtime, and Detta recognises speech on the Mac's Neural Engine.
- Does Fermion sign a BAA or data processing agreement?
- Contract terms are agreed per deployment; write to contact@fermionresearch.com with your requirements.
- How accurate is on-device speech recognition?
- Phonon-2.5 English averages 4.82 % word error on the seven Open ASR Leaderboard sets; Phonon-2.5 averages 11.71 % across 25 languages on FLEURS. Both are open, so the numbers can be re-run on your own audio.
- Can we add our own terminology?
- Yes. The runtime takes a hotword list per request, and Detta has a custom dictionary.
Related
Sources and method
- Phonon numbers: Phonon-2.5 release post (2026-10-22); throughput: Phonon-2 model page. Hotwords: /docs/speech/ (fermion-research 0.2.15).
- Detta behaviour: Detta privacy policy and the 2026-10-04 Wi-Fi-off test (lane detta_vs_wispr_20261004).
- No BAA or contract claim is made; FOUNDER_ITEMS.md.
Canonical: https://www.fermionresearch.com/deploy/on-device-speech-recognition/