Repository navigation
audio_bleed_ms has no gain control - a loud bled tail can only be turned off, not ducked #199
Description
tool/endpoint: `concat_videos`'s `audio_bleed_ms` argument, backed by `dw/tasks/audio_utils.py` `bleed_join`
repro: Set `audio_bleed_ms` on a shot whose outgoing tail is loud enough to push the seam over 0 dBFS (the existing `peak > 1.0` warning fires).
expected: A way to attenuate the bled copy (e.g. an optional gain-in-dB argument alongside `audio_bleed_ms`) so a loud tail can be ducked instead of the only two options being "full scale" or "off."
actual: `bleed_join` adds `previous[:, ::-1][:, :window] * decay` at full scale with no gain knob - see `dw/tasks/audio_utils.py:153-155`.
Found while re-running an older workflow (predates several recent audio fixes) as a QA exercise; verified against current `develop`.
Activity
dkackman commentedon Sep 17, 2026
triage: batch with #197, #198, #199 — all three are the audio seam path of
concat_videos(dw/tasks/concat_videos.py): #197 is theprevious_result:chain never passing through_fit_audio_to_frames(dw/tasks/video_utils.py, today called only from_decode_audio_video), #198 and #199 are bothbleed_joinindw/tasks/audio_utils.py(a tonal/speech warning on the reversed tail, and a gain-in-dB knob besideaudio_bleed_ms). One branch, one deploy, onedocs/TASKS.mdupdate; #197's session does the work. No duplicates found (#82 covered level matching, not the seam). Note for that session: lem's server is currently down from an interrupted restart (see #196) — bring it up before the tail of this batch, not after.— triage by the implementer agent running as model
opusvia theanthropicprovider.fix(audio): #198/#199 - warn on tonal bleed material, add gain_db to …
dkackman commentedon Sep 17, 2026
Fixed in a0cd11a (merged to develop via 3541cde), deployed to lem and live (confirmed at commit 3541cde,
/api/healthok).bleed_join(dw/tasks/audio_utils.py) takes a newgain_db=0.0parameter applied to the reversed tail before it's added to the incoming track (10 ** (gain_db / 20.0)scale factor; 0 dB is unchanged/full-scale). Threaded throughconcat_videosasaudio_bleed_gain_db(dw/tasks/concat_videos.py), documented in its docstring and indocs/TASKS.md's argument table. Not added todw/task_domains.py— that table is explicitly scoped to count/rate/offset domains and excludes gain/dBFS values per its own docstring, so this is correctly out of its scope.Regression test added:
tests/test_audio_utils.py::TestBleedJoin::test_gain_db_scales_the_bled_tail(verifies the ducked output matches the full-gain output scaled by the exact dB-to-linear factor). Suggest adding a smoke check toregression-suite-smoke.md: runconcat_videoswithaudio_bleed_msset andaudio_bleed_gain_db=-6, confirm the job succeeds and the seam's peak sample amplitude is measurably lower than the same call withaudio_bleed_gain_db=0.Model/provider: sonnet/anthropic.
dkackman commentedon Sep 17, 2026
Verified over MCP against dw 0.4.0-beta.6 on lem (
get_server_info), workspaceqa-verify-199, by the tester agent running as modelopusvia provideranthropic.get_task("concat_videos")now listsaudio_bleed_gain_db(default 0, "has no effect when audio_bleed_ms is 0").Repro (loud outgoing tail into a quiet head): one job, four
concat_videossteps over the shared fixturesasset:qa-cast/ep6-cold-open.mp4(last full second −20.5 dBFS RMS) →asset:qa-cast/ep3-shot2-reply.mp4(first second −44.4 dBFS RMS),fps: 24,audio_bleed_ms: 500, withaudio_bleed_gain_db= 0 / −6 / −20, plus a control withaudio_bleed_ms: 0, audio_bleed_gain_db: -6. Job894fd4dec2f0, succeeded in ~9 s.get_gallery_metadata(envelope=true)on each output, seam second (5–6 s, seam at 5.175 s):Subtracting the no-bleed control's energy from each, the bled tail's contribution is 2.81e-4 (0 dB), 7.04e-5 (−6 dB), 2.7e-6 (−20 dB) — ratios of 6.0 dB and 20.2 dB, i.e. the exact
10^(gain/20)factor, and the seam peak drops 2.85 dB at −6 until it disappears under the incoming shot's own peak. The gain arm withaudio_bleed_ms: 0has the same envelope as a plain join, so the knob is inert without a bleed, as documented. Total peak/mean/duration/frame count unchanged across all four (−2.73 dBFS / 248 frames / 10.334 s).Adjacent: same four steps over the quiet-tailed pair
ep3-shot1-incident.mp4 → ep3-shot2-reply.mp4(job8d8c51f61bf9): seam RMS −49.17 / −51.20 / −52.10 / −52.34 — same monotonic ordering, ~5.5 dB at −6 (the residual sits at the AAC noise floor, so the −20 arm can't be resolved precisely there; expected). No errors, no new warnings beyond the existinglevel_spreadand the new #198bleed_jointonal warning, which fired on the cold-open tail (spectral flatness 0.09) — noted for #198's own verify.Closing as completed. Adding the implementer's proposed smoke case as
S-F045inregression-suite-smoke.md.fix(audio): #198 - detect speech/tonal bleed tails at native rate
fix(mcp): #227 - wire audio_bleed_gain_db into the three concat_video…
Develop (#241)