* refactor(audio): extract capture source from recognition manager
Move audio capture out of the silence-segmented dictation loop into
vocalinux.audio.capture: a reusable PortAudioCaptureSource that owns
device resolution, format negotiation, downmixing, resampling, and
stream lifecycle (open/read/reopen/close). The dictation loop now
consumes 16 kHz mono chunks and keeps only segmentation policy (VAD,
silence counters, buffer limits, reconnection backoff).
This unblocks alternate capture sources (PipeWire system-audio capture
for #751, tdrz long-form output for #756) which need the same capture
mechanics without dictation's silence segmentation.
Behavior is identical: the same helpers run in the same order, moved
verbatim. They are re-exported from recognition_manager for existing
imports, and a CaptureSource protocol is declared in common_types.py.
Tests repoint their patches of the moved helpers to the new module.
Closes#760
* fix(audio): annotate capture API and keep a superseded session's source
* fix(audio): reset retry counters when a new capture session starts
* fix(audio): keep last-error timestamp across capture sessions
The new-session reset wiped _last_audio_error_time, so an error within five seconds of the previous session's error always took the reconnect path instead of failing fast. Only the attempt count is per-session; the timestamp is also the rapid-failure throttle.
---------
Co-authored-by: Jatin K Malik <jatinkrmalik@gmail.com>
* feat(recognition): buffer audio during model reload
* fix(recognition): recover buffered reload failures
* test(recognition): cover buffered reload edge cases
* fix(recognition): narrow buffered reload errors
* fix(settings): wait for IDLE before restoring Test Dictation callbacks
Buffered reload's stop_recognition() returns while PROCESSING. Test
Dictation must keep its text callbacks until recognition reaches IDLE
so the test utterance cannot inject into the focused app.
* style: apply Black formatting to PR 851 files
* fix(settings): cancel buffered reload before Test Dictation timeout restore
On the ~3min idle-wait timeout, do not restore live injector callbacks while
a buffered reload worker may still deliver text. Cancel via
_cancel_reload_recording off the GTK thread, then restore; if cancel is
unavailable or fails, leave test callbacks installed and show an error.
Add return/param annotations and regression coverage.
* fix(settings): keep Test Dictation waiting if buffered cancel fails
After the idle-wait timeout, if cancel is unavailable or raises, keep
polling with test callbacks and _test_active instead of marking the test
done. Ending early left the test callback as the only text destination,
so later normal dictation could land in the hidden test buffer. Restore
live injectors only once recognition reaches IDLE.
* fix(settings): reset Test Dictation timeout-cancel flag per session
Clear _test_timeout_cancel_attempted when a test starts, when entering
the buffered idle-wait, and when restoring UI so a later Test Dictation
can still cancel after timeout. Narrow the cancel-worker except clause
to AttributeError/OSError/RuntimeError/TypeError so unrelated defects
are not treated as ordinary cancel failure.
* fix(recognition): signal capture completion when buffered capture dies
A dead microphone during a buffered model reload left should_record=false
with no stop_recognition handoff, so the reload worker blocked forever on
_capture_finished.wait(). Signal the event from the audio thread's failure
teardown paths instead.
Refs VocaHQ/vocalinux#851
---------
Co-authored-by: Jatin K Malik <jatinkrmalik@gmail.com>
* fix(audio): open 3-8ch HDA mics at native layout (#813)
HDA analog capture often reports 4 channels. Opening below that native
count can abort PortAudio after read with free(): corrupted unsorted
chunks (same class of bug as #666/#673).
Try the reported channel count first when it is between 3 and 8, then
2ch and 1ch. Pulse devices that advertise 32/128 channels still open
stereo-first. Downmix N channels to mono so engines keep seeing 1ch.
Virtual-device filtering for pipewire/default (#629) is unchanged.
* fix(audio): downmix HDA capture from first channel pair
Averaging all N HDA channels halves speech when only the front pair is live.
Keep stereo mean-of-both, and for 3+ channels average channels 0 and 1.
* fix(audio): pick loudest channel for N-ch downmix
First-pair-only downmix silenced mics on ch2+. For 3+ channels, select
the loudest channel by per-buffer mean-square so speech stays at full
level whether it lands on the front or rear pair. Stereo still averages
both channels.
* fix(audio): sticky N-ch downmix channel for open stream
Pin the loudest-channel pick for the life of a capture stream so a
noise burst on another input cannot switch the mic mid-utterance.
* fix(audio): gate sticky downmix lock on speech energy
Don't pin the N-ch sticky channel from a silent/ambient first buffer;
only lock once the loudest channel clears a documented mean-square floor.
* style: isort sticky-lock import in device config tests
* feat(vad): replace amplitude-based VAD with Silero neural network
Replace the simple amplitude threshold VAD with Silero VAD, a lightweight
neural network model that classifies audio as speech/non-speech. This
addresses the root cause of #302 (silence not stopping recording) and
- Add silero_vad.py wrapper using ONNX Runtime with bundled Silero v5 model
- Accumulate 512-sample chunks with 64-sample context window per official API
- Map existing vad_sensitivity (1-5) to Silero thresholds (0.8-0.3)
- Use max() across sub-chunks to avoid missing speech onsets
- Fall back to amplitude VAD when onnxruntime is unavailable
- Add onnxruntime as optional dependency (pip install vocalinux[vad])
Closes#344
* address review: add tests, LICENSE, logging, validation fixes
- Add unit tests for SileroVAD (process, reset, fallback, sensitivity mapping)
- Add MIT LICENSE for bundled Silero VAD model in data/
- Log which VAD backend is active on init
- Validate chunk length in SileroVAD.process()
- Skip is_speech check until a full Silero chunk is processed
- Add thread safety note to SileroVAD docstring
- Document resampling rate effect on VAD resolution
- Comment explaining why load_silero_vad() catches broad Exception
* address review: bundle license, document [vad] extra, surface backend in UI
- Pack data/LICENSE into the wheel (pyproject package-data) so attribution
travels with the model
- README: short note on enabling neural VAD via vocalinux[vad]
- Settings: VAD Sensitivity row subtitle shows the active backend; if
onnxruntime is missing, an info-row beneath the group spells out the
install command and the restart requirement
- silero_vad.is_silero_available() probes onnxruntime + model presence
w/o instantiating an InferenceSession
- CI installs .[dev,vad] so Silero paths get exercised
- Integration test for _record_audio Silero branch (mocked pyaudio +
scripted SileroVAD), covering silence flush, speech reset, and
sensitivity threshold mapping at sens=1 vs sens=5
- Tests no longer skip when other modules globally mock numpy: pop the
top-level numpy ref so a fresh import gives the real package, and
swap sys.modules["numpy"] inside the _record_audio call so the lazy
import inside the function picks up real numpy too
* fix(vad): drop silence-only recording buffers
* test(vad): align stop buffer expectations
* test(vad): preserve real numpy module
* fix(install): install vad support opportunistically
* test(vad): cover recognition manager vad edges
---------
Co-authored-by: Sbenazar <127324689+Sbenazar@users.noreply.github.com>
