millet-record

Lightweight capture-only subset of millet (formerly meetscribe-record).

Records dual-channel meeting audio — your microphone on the left channel, system/remote audio on the right — into a single stereo WAV via PipeWire or PulseAudio + ffmpeg. Ships none of millet's transcription, diarization, summarization, or PDF dependencies; install footprint is ~30 MB instead of ~3 GB.

Full release history in CHANGELOG.md. Named after the Ottoman millet system. Part of the vezir ecosystem.

When to use which

Install

pip install millet-record

System deps (apt example):

sudo apt install ffmpeg pulseaudio-utils

CLI

millet check                   # verify prerequisites
millet devices                 # list audio sources
millet record                  # record dual-channel WAV; Ctrl+C to stop
millet archive                 # compress past WAV recordings to OGG/Opus

millet record writes to ~/meet-recordings/meeting-YYYYMMDD-HHMMSS/...wav unless -o is passed. See millet record --help for options.

When millet-pipeline is also installed, additional subcommands (transcribe, run, label, sync, gui, ...) become available under the same millet command via Click entry-points.

System-audio silence warning (0.5.1)

The system-audio monitor is resolved once, at session start. If the meeting app's output is later routed elsewhere — switching apps mid-call, plugging in headphones — that monitor goes silent while your mic keeps the stereo file growing. The remote side is lost with no process failure and nothing in the output to suggest it.

millet record now samples per-channel RMS while recording and warns inline the first time the system channel goes quiet:

⚠ System audio silent — remote participants may not be recorded
✔ System audio restored

session.json records system_ever_active and system_silent_detected, so a downstream tool can flag a recording that captured only one side. An unreadable sample counts as unknown, never as silent, so a transient ffmpeg hiccup cannot produce a false alarm.

Crash resilience (0.6.0)

Recorders run start_new_session=True-detached on purpose: if the controlling app (TUI, CLI) crashes, the recording survives — the meeting is still happening. 0.6.0 adds the bookkeeping to make that survivable in practice:

• recording.lock in the recordings root blocks a second concurrent recording (RecordingInProgressError, PID-checked, stale locks auto-reclaimed; MEET_RECORD_LOCK=0 disables).
• <stem>.recorder.json identifies the live recorder process (pid + owner pid + start ticks); removed on clean stop/pause.
• <stem>.session.json is now written at recording start (status: "recording") and rewritten at stop (status: "stopped").
• millet_record.capture.find_interrupted_sessions(root) scans a recordings dir for sessions that never finished, distinguishing in progress / orphaned recorder / interrupted; and millet_record.capture.recover_session(dir) stitches leftover chunks into the final WAV (repairing SIGKILL-damaged headers).

Legacy meet command (removed in 0.6.0)

The pre-rename meet console script and the meet_record import alias were removed in millet-record 0.6.0, after the announced two-minor-version deprecation window. Use millet / millet_record.

Architecture

millet-record exposes a stable package millet_record containing:

• millet_record.capture — ffmpeg-backed dual-channel capture (RecordingSession, watchdog, drain buffer)
• millet_record.audio — stereo channel reading, per-channel RMS sampling (system-silence detection), ffmpeg-based audio compression
• millet_record.utils — formatting helpers (HH:MM:SS, file sizes)
• millet_record.languages — language constants used by capture flow
• millet_record.cli — millet console-script entry point

millet-pipeline depends on this package and re-uses these modules, plus its own heavy modules (transcribe, label, voiceprint, summarize, sync, pdf, gui).

macOS (Apple Silicon)

pip install millet-record on macOS 14.4+ Apple Silicon ships a bundled meet-record-mac Swift sidecar that captures via Core Audio Process Tap + AVAudioEngine — no PulseAudio, no BlackHole, no extra install. millet record uses it by default.

Note: the Swift binary itself is still named meet-record-mac for now — renaming would require macOS code-signing bundle-path changes that aren't worth doing as part of the package rename. Tracked as a follow-up; doesn't affect end users.

First run prompts for Microphone and System Audio Recording permissions via the standard macOS TCC dialogs; both are required for full dual- channel capture (mic on left, system on right). See mac/README.md for the sidecar's CLI surface, level analysis recipes, and environment variables.

On macOS Sequoia 15+, Apple removed the manual + button from System Settings → Privacy → Microphone, so users can no longer add permissions before running the app. Instead, run millet check: it invokes the sidecar's request-permissions routine, which calls AVCaptureDevice.requestAccess(for: .audio) to trigger the TCC dialog, and reports which permission is still missing. (request-permissions is a subcommand of the bundled meet-record-mac binary, not of the millet CLI — millet check is the user-facing entry point.)

Set MEET_RECORD_MAC=0 to force the legacy ffmpeg+PulseAudio path (diagnostic kill switch only — that path will fail on a stock macOS install because there is no PulseAudio device). Intel Macs and macOS < 14.4 are unsupported.

License

GPL-3.0-or-later, same as parent millet.