| name | ambient-faster-whisper-stt |
| description | Local faster-whisper STT provider for Ambient speech input validation. |
Use this package only through Ambient's STT provider path. It implements:
faster_whisper_transcribe --audio <wav> --language <language> --output-json <path>
Operational notes:
- The default path is intentionally small:
faster-whisper==1.1.1, tiny.en, CPU, int8, beam size 1.
- This package is adapter-only. It bundles Ambient descriptor metadata, these instructions, the Node wrapper, the Python runner, and tests; it does not bundle
uv, Python, faster-whisper wheels, CTranslate2 binaries, or model weights.
requests is installed explicitly in the uv environment because the live macOS smoke found the direct faster-whisper import path needed it.
- First real transcription may populate local uv/Python/faster-whisper/model caches. Treat health output as runtime-readiness guidance, not proof that all model assets are already downloaded.
- When validating first-run behavior, use
pnpm run stt:faster-whisper-clean-cache from the repo root so uv, Python, Hugging Face, and XDG caches are isolated in the validation artifact.
- This is setup/control evidence for the Ambient STT product path. Do not present
tiny.en as multilingual quality evidence.
- Use a larger model only after setting
AMBIENT_FASTER_WHISPER_MODEL intentionally and recording the runtime/quality tradeoff.