The single encode at save time was %mp3(bitrate=16), and the comment said why:
it matched what Liquidsoap pushes to Icecast. That was the wrong thing to
match. Icecast is the LISTENING path and 16 kbps is a bandwidth budget for a
live stream; this file is the ACCURACY path -- it is what Whisper transcribes,
and CLAUDE.md is explicit that everything downstream is hostage to it. P25 has
already been through a vocoder, so 16 kbps MP3 stacked a second lossy stage on
the one copy that had to stay faithful.
FLAC instead. Lossless, so the bytes Whisper receives are the bytes PulseAudio
captured. ~1.3 MB/min against 120 KB/min, which keeps a 600 s call (the time
cap) around 13 MB -- inside Whisper's 25 MB request cap and well inside
upload_max_bytes. Icecast's own 16 kbps stream is untouched; nothing about
live listening changes.
Capture, buffering, silence detection and the byte-offset trim are all
unchanged: they operate on raw PCM and never saw the encode. The sample rate
stays pinned to pcm.SAMPLE_RATE so the encode remains a straight pass -- the
trim arithmetic depends on that, and Whisper resamples to 16 kHz itself.
encode_mp3 is now encode_recording, the upload sends audio/flac, and the test
that pinned the old contract now pins losslessness instead, including an
assertion that no bitrate constant comes back.
This is the before/after boundary for STT quality. Last night's window is the
16 kbps baseline.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two unauthenticated surfaces closed on the edge node.
Dashboard and API: the local dashboard and every /api/* route were open to
anything on the node's LAN. Adds a login page plus session-cookie auth for
the browser, and cookie-or-Basic for the API so scripted callers stay
possible. Passwords are hashed with stdlib scrypt (no new dependency, this
runs on a Pi) and compared in constant time; the salt and session-signing
secret persist in credentials.json. Startup warns while the default password
is still in place. No non-browser callers of the node API exist today
(C2 talks to nodes over MQTT and nodes call C2 outbound), so nothing breaks.
Adds python-multipart, which FastAPI's Form() needs for the login POST and
which was missing from requirements entirely.
MQTT: nodes authenticated with a shared drb-node password, and the broker
ACL keyed off %c — the client-supplied client id — so any holder of that one
password could claim another node's topic namespace. Nodes now connect as
username=<node_id>, password=<their C2-issued api_key>, which mosquitto's
dynamic-security plugin checks, with the ACL keyed off the authenticated %u.
TLS is gated on MQTT_TLS and uses default CA verification.
The old key_request MQTT path stays in place behind TODO(mqtt-cutover)
markers as the fallback until the cutover is proven; a node with no api_key
on disk logs a clear repeated refusal rather than spinning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The control channel was wrong in both directions. Grants fire 0.84-1.62s
before anyone speaks, and srcaddr can drop to 0 while someone is still
talking - one recording came back "-1.61s lead, -0.00s tail", the trim
finding nothing to remove because the window had closed on live speech.
Confirmed by ear: the cut lands at a word boundary on an unfinished word.
Audio is ground truth for WHEN. The console remains the only source of
WHO, so it still supplies talkgroup, alias and rid.
START voice onset in the captured audio, with a 0.25s pre-roll that
now covers only chunk quantisation and threshold ramp-up rather
than a variable control-channel offset.
STOP call_silence_timeout seconds of silence heard in the audio.
LABEL resolved AT CLOSE from a bounded rolling history of console
observations overlapping the window, +4s/-2s, because there is
no guaranteed ordering between a grant and its audio.
SPLIT a console talkgroup change still forces a cut, since two calls
with no silence between them would otherwise merge into one.
Capture now emits raw PCM instead of MP3. Silence detection becomes
integer arithmetic per chunk with no decode, trimming becomes a byte
offset slice rather than a second ffmpeg pass, and MP3 encoding happens
exactly once at save - uploads are no longer double-encoded.
Audio with no talkgroup anywhere in its window is discarded rather than
uploaded: an untagged call silently poisons incident correlation, which
is worse than losing the audio. Logged at ERROR and counted on
/api/status.
When capture produces no audio at all the old console state machine
still runs, so a node with a broken audio path keeps reporting radio
activity. That is now the only consumer of call_idle_timeout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three unrelated failures from `make test` on the node:
test_metadata_watcher: two tests asserted the old zero-pad tgid_change
contract. Both were off by exactly call_tail_pad_seconds, i.e. the code
was doing what f1de157 intended and the tests encoded the behaviour we
deliberately changed. Updated to expect the pad.
test_pulse: four async tests errored with "async def functions are not
natively supported". Root cause is not the tests - pytest.ini sets
asyncio_mode = auto but the Dockerfile only copies app/ and tests/, so
the container had no pytest config at all and fell back to strict mode.
That is also why these passed in a local venv and failed on the node.
Copy pytest.ini into the image, and add explicit @pytest.mark.asyncio so
the tests hold up regardless of how pytest is configured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The recording window is anchored to OP25 control-channel timestamps, but
the buffered audio lags those by roughly 1.5s. Trim logs across seven
calls measured the offset at 0.84-1.62s, consistently present.
At a 1.0s pad a short call closed its window before the voice arrived:
a 0.97s control-channel call closed at T+1.97 while voice started around
T+1.5, capturing ~0.4s of speech and cutting mid-word. Confirmed by a
0.57s file whose final 0.10s measured -12.2dB against its own -18.2dB
average - clipped speech, not a tail - and by two short calls that
logged no trim at all because no trailing silence remained.
Being generous is free here: trim_silence already strips trailing
silence back to the guard margin before upload, so long calls are
unaffected while short ones gain the window they need. Over-capture
costs nothing; under-capture loses words permanently.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The pulse_socket named volume survives container recreation, so after a
compose recreate the previous container's /run/pulse/pid and native
socket were still present. PulseAudio read the stale pid file, decided a
daemon was already running, and refused to start:
E: [pulseaudio] pid.c: Daemon already running.
The entrypoint still reported "PulseAudio socket ready" because it only
checked that the socket file existed - and a stale one did. Capture then
failed in a restart loop against a dead daemon.
Readiness in both the op25 entrypoint and drb-edge-node now means a
pactl probe actually succeeds. Stale pid/socket are removed only when
that probe fails, so a live daemon's socket is never deleted.
pulseaudio-utils was missing from the edge-node image (only libpulse0
was installed), so no pactl binary existed there at all - added.
Capture exits are now classified: a missing source logs at ERROR and
names the configured PULSE_SOURCE, rather than looking identical to
"daemon not up yet". Retrying forever against a wrong source name is how
the April PulseAudio failure stayed hidden.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Measured six recordings off a live P25 node and found two independent
defects causing lost audio at the end of calls:
- stop_recording() sliced the ring buffer immediately, so if the MP3
muxer had not yet delivered the tail the file was silently short.
Now waits (bounded, 2s) until buffered audio covers the end epoch.
- TGID-change closes used the new grant's epoch as the end with no pad
at all, guaranteeing truncation on every split. Tail pad is now a
setting, default raised 0.5s -> 1.0s.
The ring buffer also capped maximum call length: a call longer than the
buffer had its front silently clamped. The ring now serves the pre-roll
only, with a per-call accumulator for the rest, bounded at 4.8MB.
Clamping is loudly warned rather than silent.
Uploads averaged 63% silence, which inflates STT cost and is a known
Whisper hallucination trigger. Leading/trailing silence is now trimmed
conservatively (-40dB, 0.25s guard, internal pauses untouched).
started_at/ended_at still describe the call; new audio_* fields carry
the trimmed audio bounds so playback can map back to wall clock.
All-silence recordings are skipped and logged instead of uploaded.
Also: log measured control-channel idle on idle-timeout closes so
CALL_IDLE_TIMEOUT can be tuned from data, and quiet the httpx logger
which emitted ~170k lines/day of poll noise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Logan CusanoandClaude Opus 5