Repository navigation
Pluggable speech-to-text endpoint for restricted / air-gapped networks #2507
t79699271949
started this conversation in
Ideas
Replies: 2 comments
|
No please fork |
0 replies
|
Fair enough — I'll keep it in the fork. If someone else runs into the same wall, the code and a ready portable build are here:
It's batch only — press, speak, release. Tested against a whisper.cpp server from a machine with no local model, with the language hint passed through. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The problem
Handy assumes speech recognition always happens on the local machine. In
companies whose security team forbids downloading models and blocks arbitrary
outbound traffic, that single assumption makes the app impossible to approve —
even though everything else fits: local-first, MIT, no telemetry.
The organisation already runs an internal speech-to-text server that speaks the
OpenAI API (vLLM, whisper.cpp
server, LM Studio all do). Handy can't use it,because speech recognition has no configurable backend.
There is a second blocker behind the first: recording cannot even start without
a downloaded local model. With no way to download one, the microphone never
opens. So it isn't only "no option to point at a server" — the air-gapped path
is closed twice.
Why this is a natural extension, not a new subsystem
Post-processing already works exactly this way: a
customprovider with aneditable
base_urland an OpenAI-compatible request. Speech recognition is theone place where the same pattern is missing. The change follows an existing,
accepted design.
Proposal
One new setting block, off by default:
enabled— when off, nothing changes at all;base_url— e.g.http://stt.internal:8000/v1(path/audio/transcriptionsis appended);model— whatever the server expects (whisper-1,large-v3, …);api_key— optional Bearer token;timeout_secs;languagehint.When enabled, the recorded buffer is sent as
multipart/form-datato/audio/transcriptions, and the returned text follows the normal path(post-processing → paste → history). Batch only: press → speak → release.
The local model gate is skipped while the mode is on, so a machine with no
downloaded model can still dictate against the internal server.
What this is not
separately if there's interest.
Scope
A settings struct, one module that wraps the PCM buffer into WAV and posts it,
one branch where transcription starts, one settings block in Experimental, and
unit tests (including a local-socket test that asserts the WAV body and the
model field actually go out).
Question for maintainers
Would a PR of this shape be considered, given the feature freeze — or is the
external-backend idea out of scope for Handy by design?
All reactions