Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

Offline speech recognition in Python

Every library here downloads a model once and then transcribes on your own machine. They differ in model accuracy, languages, size and what they need installed besides pip.

Speech quickstart

The short answer

To transcribe audio offline in Python, pip-install a library that runs an open model locally: fermion for the Phonon models (pip install fermion-research), openai-whisper or faster-whisper for Whisper, Vosk for small per-language models, or Moonshine for streaming on small devices. Each downloads its model on first use and then runs without a cloud service.

For English, Phonon-2.5 English averages 4.82 % word error on the seven Open ASR Leaderboard English sets from a 158 MB model; for 25 European languages, Phonon-2.5 is 156 MB. Whisper covers 99 languages from 75 MiB (tiny) to about 3 GB (large).

Python libraries compared

Install commands and requirements from each project's own documentation, read 2026-11-13.

LibraryInstallModelsLanguagesAlso needs
fermionpip install fermion-researchPhonon-2.5, Phonon-2.5 English, Phonon-2 (156 to 164 MB)25 European, or EnglishPython 3.10 to 3.14; MLX stack on Apple silicon, torch (CPU) on Linux and Windows
openai-whisperpip install -U openai-whisperWhisper tiny to large, turbo99ffmpeg
faster-whisperpip install faster-whisperWhisper on CTranslate299Python 3.9+; no ffmpeg (PyAV); cuBLAS and cuDNN for GPU
Voskpip3 install voskabout 50 MB per language20+a downloaded language model
Moonshinepip install moonshine-voicestreaming models from about 1 MB10 for speech-to-textnothing listed
mlx-whisperpip install mlx-whisperWhisper on MLX99Apple silicon, ffmpeg

Transcribe a file with fermion

The command line prints the transcript to standard output; --json adds word timestamps. transcribe reads wav, flac, ogg and aiff; for mp3 and m4a it prints the ffmpeg command to convert first.

pip install fermion-research
fermion transcribe phonon-2.5 meeting.wav
fermion transcribe phonon-2.5-en call.wav
fermion transcribe phonon-2.5 clip.flac --json

# On Linux, install the CPU build of torch first to skip the larger GPU build
pip install --no-deps torch --index-url https://download.pytorch.org/whl/cpu
pip install fermion-research

From Python code

The Python API loads a speech model once and transcribes files or arrays; this example is from the documentation and uses Phonon-2 with hotwords:

import fermion

speech = fermion.load_speech("phonon-2")
text, decode_s, audio_s = speech.transcribe("call.wav", hotwords=["Ada", "Quillon", "Okonkwo"])
print(text)

Running on a machine with no internet

The model downloads once into the Hugging Face cache (move it with HF_HOME or FERMION_CACHE_DIR). For a machine with no network at all, fetch the model on a connected machine with the transcribe command's --download-only flag, which fetches and verifies the model, prints its directory and stops; copy that directory across and pass it as the model argument, since every speech command accepts a local model directory.

To serve other programs on the same machine, fermion serve exposes POST /v1/audio/transcriptions on localhost.

fermion transcribe /path/to/model-directory meeting.wav
fermion serve phonon-2.5

Questions

How do I do speech recognition in Python without the internet?
Install a library that runs an open model locally, such as fermion (pip install fermion-research), openai-whisper, faster-whisper or Vosk. After the model downloads once, transcription runs on your machine.
Which offline Python speech library is most accurate?
It depends on the model it runs. Phonon-2.5 English averages 4.82 % word error on the seven Open ASR Leaderboard English sets; Whisper large-v3, available through openai-whisper and faster-whisper, averages 6.04 % on the same sets in the 2026-09-25 snapshot.
Do I need a GPU for offline speech recognition in Python?
No. fermion runs the Phonon models on plain CPUs on Linux, Windows and macOS, and faster-whisper and Vosk run on the CPU. A GPU speeds up the larger Whisper models.
Does offline speech recognition need ffmpeg?
openai-whisper and mlx-whisper need ffmpeg. faster-whisper decodes audio with PyAV instead, and fermion reads wav, flac, ogg and aiff directly.
Can I get word timestamps offline?
Yes. fermion transcribe --json returns the start and end of every word for the Phonon-2 and Phonon-2.5 models, and faster-whisper and mlx-whisper take word_timestamps=True.
Which Python versions does fermion support?
Python 3.10 through 3.14 on macOS, Linux and Windows.

Related

Sources and method

  • fermion commands, requirements, cache and local-directory behaviour: /docs/quickstart/, /docs/speech/ and /docs/cli-reference/ (fermion-research package documentation), read 2026-11-13.
  • openai-whisper: github.com/openai/whisper; faster-whisper: github.com/SYSTRAN/faster-whisper; mlx-whisper: github.com/ml-explore/mlx-examples; Vosk: alphacephei.com/vosk; Moonshine: github.com/moonshine-ai/moonshine; read 2026-11-13.
  • Phonon English averages: Fermion Research measurement with the Open ASR Leaderboard's scoring code on the full seven test sets (Phonon-2.5 release post, 2026-10-22).
  • Rival English averages: Open ASR Leaderboard, seven English sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli), snapshot 2026-09-25.

Canonical: https://www.fermionresearch.com/speech/offline-speech-recognition-python/