You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Agent] Implement the voice-enhancement audio filter (#395) #652
The per-stream "voice enhancement" option (discussion #395) is stored and passed around, but does nothing:
src/video/mp4_writer_utils.c:62-70: voice_enhancement_enabled, filter_graph, filter_buffersrc_ctx and filter_buffersink_ctx exist, with the comment "the actual filter chain (afftdn / highpass / lowpass ...) is deferred to a follow-up PR — for now this is a stub".
The setting is stored in the DB (db_streams.c), exposed through the streams API (api_handlers_streams_get.c / _modify.c) and shown in web/js/components/preact/StreamsView.jsx. The help text says it "has no effect yet".
libavfilter isn't linked anywhere today.
Expected behavior and scope
Link libavfilter, using a CMake option such as ENABLE_AUDIO_FILTERS that defaults ON when found. Add it to the Dockerfile / Debian packaging deps.
In transcode_audio_packet(), once the decoder frame format is known, lazily build the graph: abuffer → highpass=f=100 → lowpass=f=8000 → afftdn → aformat(<encoder sample fmt/rate/layout>) → abuffersink. Tune the defaults for speech. Push decoded frames through it before encoding, and free the graph in transcoder cleanup.
This only applies on the path where audio is already being transcoded. Document that passthrough (copy) audio isn't filtered, or force transcode when enhancement is on, and say which.
If the graph fails to build, log once and continue unfiltered. Recording must never fail because of this.
Update the UI help text and translations (web/public/locales/en.json) so it no longer says "no effect yet".
Acceptance criteria
With enhancement on, recorded audio is audibly filtered. A test can check this by comparing spectral energy below 100 Hz before and after.
With it off, the output is byte-for-byte the same pipeline as today.
A build without libavfilter compiles, and the option logs that it's unavailable.
No leaks under ASan (sanitizer workflow).
Validation
Unit test feeding generated PCM (sine plus low-frequency hum) through the transcoder with enhancement on, then asserting the hum is attenuated.
Run the full ctest after building all targets, plus the sanitizer CI.
Constraints and rollback
Watch the CPU cost on embedded profiles. afftdn is moderately expensive; consider leaving it out of the embedded profile.
Rollback: revert, and the opt-in goes back to being a no-op.
Activity
Merge pull request #671 from opensensor/feat/voice-enhancement-filter…