Skip to content

speech.md

Latest commit

 

History

History
64 lines (49 loc) · 2.21 KB

speech.md

File metadata and controls

64 lines (49 loc) · 2.21 KB
Raw

Local speech-to-text

sudo sibilla speech enable puts a transcription endpoint on the same gateway, behind the same keys, as everything else this machine serves. The engine is whisper.cpp, repackaged from its pinned upstream release exactly like llmfit (recorded sha256, no installer scripts), and the Whisper model arrives like every other catalog artifact: verified against the digest in the GPG-signed catalog, refused on mismatch, with no override. Audio is transcribed on this machine and goes nowhere else.

Use

$ sudo sibilla speech enable
speech model verified: sha256:60ed5b...
speech enabled at /v1/audio/transcriptions (same API keys as the API)

Then call the OpenAI transcription shape:

$ curl http://YOUR_HOST:8080/v1/audio/transcriptions \
    -H "Authorization: Bearer YOUR_KEY" \
    -F file=@recording.wav \
    -F response_format=json
{"text": " ..."}

Any OpenAI SDK works the same way (client.audio.transcriptions.create), pointed at this machine with sibilla connect --env.

response_format accepts the upstream server's formats (json, text, srt, vtt, verbose_json), which covers subtitles directly.

Formats and ffmpeg

Without ffmpeg installed the server accepts 16 kHz mono WAV only, and sibilla speech enable says so. With ffmpeg (sudo apt install ffmpeg, then enable again) most audio formats convert on the fly. ffmpeg is a Recommends, not a hard dependency: the appliance stays minimal by default.

Configuration as code

SPEECH=on|off is a profile key like the other toggles: sibilla apply converges it, apply export records it, apply check reports its drift.

Air-gapped delivery

Put the model file at /var/lib/llmd/models/whisper/ggml-base.bin by any offline means; sibilla speech enable verifies the existing file against the catalog digest and keeps it, downloading nothing. A file that does not match is refused, exactly like a network download.

Model

One model ships in the catalog: Whisper base (multilingual, ~148 MB, MIT), a deliberate default that transcribes well on CPU. The catalog can grow other sizes the day someone brings a use case that base does not cover.