Fermion ResearchFermion Research

Phonon/facts checked 13 November 2026

Self-host a speech-to-text server with an OpenAI-compatible endpoint

fermion serve turns any Phonon model into a local transcription server with the same route shape as OpenAI's audio endpoint. No GPU required.

Speech recognition docs

One command

Install the package and serve a model. The audio routes mount in place of the chat routes.

pip install fermion-research
fermion serve phonon-2.5
# Listening on http://127.0.0.1:8000

curl -s http://127.0.0.1:8000/v1/audio/transcriptions -F file=@clip.wav
# {"text": "The transcript."}

Routes

  • POST /v1/audio/transcriptions: one-shot file transcription; longer audio is handled in chunks.
  • GET /v1/audio/stream: live transcription over a standard WebSocket (see the streaming docs).
  • Health endpoint stays unauthenticated; every /v1 route takes a bearer token when --api-key is set.

Beyond localhost

A non-loopback bind requires an API key; without one the server refuses to start.

fermion serve phonon-2.5 --host 0.0.0.0 --port 8000 --api-key "$FERMION_API_KEY"

Docker

Two images serve the same endpoint as fermion serve, so a client written against one works against the other.

docker run --rm -p 8000:8000 -v phonon-cache:/home/phonon/.cache ghcr.io/fermionresearch/phonon-cpu:2.0.12 serve phonon-2.5
docker run --rm --gpus all -p 8000:8000 -v phonon-cache:/home/phonon/.cache ghcr.io/fermionresearch/phonon-cuda:1.0.11 serve phonon-2.5

Which model

  • phonon-2.5: 25 languages, 156 MB, 11.71 % average FLEURS word error.
  • phonon-2.5-en: English only, 158 MB, 4.82 % on the seven Open ASR Leaderboard sets.
  • phonon-2: the earlier English model, 164 MB, 5.21 %; the one with a Core ML package for the Apple Neural Engine.

Throughput

Published Phonon-2 figures, which Phonon-2.5 shares in 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. An eight-core CPU box therefore clears more than a hundred hours of audio per hour.

Questions

Is the endpoint compatible with the OpenAI SDK?
The route is POST /v1/audio/transcriptions with a multipart file field and a JSON text reply, the same shape OpenAI clients post to. Point the client's base URL at the server.
Do I need a GPU?
No. The CPU image and the pip package run on x86 and Apple silicon; the CUDA image adds NVIDIA GPUs.
Can it transcribe in real time?
Yes, GET /v1/audio/stream carries live transcription over a WebSocket.
How do I protect the server?
Pass --api-key; it is required for any bind other than localhost, and every /v1 route then needs the bearer token.
Which languages can the server transcribe?
With phonon-2.5, 25 European languages; with phonon-2.5-en or phonon-2, English.

Related

Sources and method

  • Commands, routes and flags: /docs/speech/ and /docs/serve/ (fermion-research 0.2.15, 2026-10-24).
  • Throughput: Phonon-2 model page, published figures.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/self-host/