Files
faceclaw/docs/models/whisper-small-fr/README.md

3.9 KiB
Raw Permalink Blame History

On-device Whisper French

Android Settings → Voice now offers On-device (Whisper French) and a separate On-device model: Whisper (French, small) download (~375 MB decimal). Download first, then choose the provider. Recognition stays on the phone; no API key is required. This is speech-to-text in French, not translation.

This provider always requests language=fr, task=transcribe, without writing the shared language preference. Switching back to existing Moonshine, English Whisper or multilingual Whisper restores their previous behavior. Downloads, cancellation/resume and deletion reuse the existing model manager and leave other model directories untouched. Missing French files do not silently select multilingual base. iOS and the Microphones app's separate Moonshine caption engine are outside this addition.

Provenance and distribution

  • Source: https://huggingface.co/bofenghuang/whisper-small-cv11-french
  • Pinned revision: 83d7df2928fb2218721c543dcc0be29ef8db2f96.
  • The model author's declared license is Apache-2.0; LICENSE.txt is distributed with the downloadable model. The original Whisper implementation is MIT; sherpa-onnx's exporter is Apache-2.0.
  • Exporter: sherpa-onnx v1.13.0 scripts/whisper/export-onnx.py.
  • Export: strict tensor-name conversion from HF Whisper to OpenAI Whisper, then upstream sherpa tensor-cache export, opset 17, dynamic int8 MatMul.
  • Assets: https://git.undefined-dev.fr/Blendman974/faceclaw/releases/tag/models-whisper-small-fr-v1
  • Model files use the public Gitea generic package registry (faceclaw-asr, version whisper-small-fr-v1), because the encoder and decoder exceed the instance’s 100 MB attachment limit. No infrastructure limit is changed.
  • The JSON manifest records file lengths and SHA-256 hashes. Both the source checkpoint and exporter are checksum-verified before execution. The existing Android downloader checks each downloaded asset against its pinned hash.
  • Only encoder, decoder, vocabulary and license are downloaded to the phone; the original checkpoint and unquantized intermediate models are not shipped.

Reproducing the export

Use an isolated Python environment with the versions used for this release: Python 3.14, torch 2.9.0+cpu (CPU wheel index), openai-whisper 20250625, onnx 1.23.2, onnxruntime 1.31.1 and sherpa-onnx 1.13.0. These are development requirements only; no Python or additional runtime is added to the app. On NixOS expose the host C++ runtime and zlib through LD_LIBRARY_PATH.

python tools/export-whisper-small-fr.py --output /tmp/whisper-small-fr
python tools/test-whisper-small-fr.py \
  --model-dir /tmp/whisper-small-fr --wav /path/to/french-test.wav

The exporter downloads the pinned source (~484 MB), writes unquantized and quantized intermediate models, and produces manifest.json. Artifact changes require a new versioned release and explicit catalog hash changes; never replace v1 assets in place. These scripts do not upload or publish files automatically.

Validation

The exported int8 files were actually decoded on CPU with sherpa-onnx 1.13.0, the same version used by Android. A public 4.97-second French sample from csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3/test_wavs/fr.wav produced:

Ne vous demandez pas ce que votre pays peut faire pour vous, demandez-vous plutôt ce que vous pouvez faire pour lui.

This is a smoke test, not a recognition-quality benchmark or a measurement of Pixel latency. Application tests cover provider persistence, French selection without rewriting other languages, switching back, capture sharing, interrupted download/resume, file hashes and independent deletion. Kotlin tests cover the native file selection and preservation of existing filenames/defaults.

Hardware checks remain pending: real French dictation through phone and glasses microphones, latency and RAM on the Pixel, airplane-mode use after downloading, reconnection and long utterances. No firmware changes are required.