Multi-track voice memos for macOS. Record two people wearing two lapel mics, get back two audio files and two transcripts — one per speaker.
Voice Memos records one track. If you clip a DJI Mic Mini transmitter to each person and plug the receiver into your Mac, the hardware is already keeping them apart — TX1 on the left channel, TX2 on the right. Lapel is the app that stops throwing that away.
A two-person conversation recorded to one mixed track is a transcript full of guesses about who said what. Diarization models exist and they are mediocre. But if each speaker is already wearing their own microphone, the separation is physical — perfect, deterministic, and free. Lapel just refuses to discard it.
- Sees the receiver arrive. Hotplug detection via CoreAudio, no polling.
- Tells you how many lapels are actually live — not how many the receiver could take, how many are switched on and transmitting right now.
- Meters each transmitter separately, with proper ballistics and clip detection.
- Records to one file per speaker, named by whoever was wearing that mic. Names are remembered for next time, and a recording made before anyone was named can be attributed afterwards — click the name, type, press Return, and the transcript follows.
- Transcribes each track on-device, so the transcript is already attributed.
- Mutes mic bleed before transcribing. Two lapels at one table each hear both people; Lapel gates each track on which mic was loudest, so a line lands only on whoever said it — one transcript, no duplicates.
- Warns you when the receiver is in Mono mode, where the two lapels are mixed together and separation is impossible — with the fix, on the hardware, in words.
- Exports a shareable pair: all speakers mixed into one audio file, and one combined transcript with every line attributed — as plain text, Markdown or SRT.
- Trims before you export, against a waveform of the combined mix, with playback of exactly the stretch you selected.
Under construction. Working today:
| ✅ | Device enumeration, hotplug and channel-mode detection |
| ✅ | Level metering, live transmitter count, speaking indicator |
| ✅ | Recording state machine, per-channel file writing, session store |
| ✅ | lapel-probe terminal harness |
| ✅ | SwiftUI app — live meters, transport, session library |
| ✅ | Combined audio + transcript export (txt, Markdown, SRT) |
| ✅ | On-device transcription with SpeechAnalyzer (requires macOS 26) |
| ✅ | Crosstalk gate — one attributed transcript from two mics in one room |
No app needed — the probe verifies everything below the UI against your own gear:
swift run lapel-probeInput devices
────────────────────────────────────────────────────────────────────────
▸ DJI MIC MINI 2 channels 48000 Hz usb
MacBook Air Microphone 1 channel 48000 Hz builtIn
────────────────────────────────────────────────────────────────────────
DJI MIC MINI — 2 of 2 microphones live
TX1 [███████████████|·················] -18.4 dBFS ● speaking
TX2 [████|···························] -46.1 dBFS ○ live
Plug and unplug the receiver, or press its mode button, while it runs.
No receiver to hand? Write a demo session and open the app:
swift run lapel-probe --demo-session ~/Library/Containers/com.4mn.lapel/Data/Library/Application\ Support/Lapel/SessionsAlready recorded something? The same pipeline the app runs, with every stage printed:
swift run lapel-probe --transcribe-check <session folder> # gated, as the app does it
swift run lapel-probe --transcribe-check <session folder> --raw # no gate, for comparison
swift run lapel-probe --transcribe-check <session folder> --save # write transcript.json beside the audio-
macOS 14 or later. On-device transcription requires macOS 26 and a build made with Xcode 26 — on older toolchains the engine compiles out and the app says so.
-
A DJI Mic Mini, DJI Mic 2, or any multi-channel USB audio input
-
The receiver must be in S (Stereo) mode. In M (Mono) it mixes both lapels together before your Mac ever sees them. Lapel will tell you if it is — including the case the channel count cannot reveal, below.
On a DJI Mic Mini, double-click the pairing button to switch between Mono and Stereo. There is no labelled mode switch and no on-screen indication, so this is easy to miss. Verified on hardware: stereo separation does apply to the USB-C output, not just the 3.5 mm camera jack.
The rule the codebase is built around: hardware types stop at the bridge.
Metering, receiver detection, transmitter presence, the recording state machine and
the session store contain no CoreAudio or AVFoundation types whatsoever. Audio
reaches them as [[Float]] — one array per channel. Everything above the bridge is
therefore testable against synthetic signals with exact arithmetic, and the entire
suite runs on a CI machine with no audio interface attached.
Exactly three files talk to hardware:
CoreAudioDeviceEnumerator— reads the device list from the HALAudioDeviceMonitor— hotplug and channel-mode listenersAudioCapture— the AVAudioEngine tap, deinterleaving to[[Float]]
The combined mix is drawn as a waveform you can drag the ends of, scrub, and play. Playback stops at the out point, so what you audition is exactly what export contains.
Trimming is non-destructive — the per-speaker recordings are never modified — and the transcript is re-timed alongside the audio, so the two exported files still agree. A turn straddling a cut is clipped rather than dropped, because someone mid-sentence at the boundary still said something inside the clip.
The per-speaker files are the point, but they are awkward to send to someone. So export produces the shareable pair alongside them: every track mixed to one audio file, and one transcript with each line attributed.
Export runs off the main thread with a progress bar and a working Cancel; a
cancelled export removes the file it was writing. Audio is handed to the encoder in
blocks of about a second. That matters more than it sounds: passing a 45-minute mix
to the AAC encoder as one buffer had it grinding for hours, and the same file in
blocks encodes in six seconds. swift run lapel-probe --export-check <session> <out>
times it.
The mixdown is a sum, not an average. Averaging is the textbook answer and it is wrong here — two people taking turns is the normal case, so only one track is loud at any instant, and dividing by the track count would halve the volume of a recording that never came close to clipping. Instead the sum is taken at full weight, and a single gain is applied across the whole mix only if its peak would exceed −1 dBFS.
Mixing the audio discards the separation, so the transcript has to carry it in the
text. Names you typed are used as written; an unnamed track becomes Speaker 1
rather than TX1, because a hardware label has no business in a document meant to
be read. A numbered mode overrides names entirely, for sharing a transcript without
naming anyone.
Separate tracks settle whose microphone, not whose voice. Two people wearing lapels at the same table are each picked up by both transmitters — a hand's breadth from their own mouth, a table's width from the other's — and a transcriber given the raw tracks dutifully returns every sentence twice, once per mic.
But quieter is the whole answer. Measured on a real 45-minute conversation, the speaker's own mic wins by 6–15 dB and fewer than 4% of speech windows are within 3 dB of a tie. So before transcribing, Lapel compares the tracks in 20 ms windows, decides who owned each one, and mutes every other track there — with a short ramp either side of each cut, a hold so a breath does not hand the floor over, and a blip filter so a click on the wrong mic does not count as a turn. The engine then only ever hears a person through their own microphone. Any line it still produces where the other mic owned the room is thrown out afterwards.
The recordings on disk are never touched; the gated copies are temporary. Genuine overlap — both people talking, each mic owning part of the stretch — is kept on both, because both of them really did say something. It can be switched off for mics that could not hear each other, such as a call.
A real DJI Mic Mini enumerates as Wireless Mic Rx, two input channels, 48 kHz,
USB — even in Mono mode, where it sends the identical mix down both channels. A
receiver can therefore claim stereo and still be incapable of separating speakers.
So Lapel compares the two channels sample for sample. If they match while carrying signal, it says so and names the fix. Silence is never treated as evidence: two channels of digital zero are identical, but that is exactly what a correctly configured stereo receiver looks like with both lapels switched off.
The device's product name also contains no vendor string at all, which is why detection matches on the manufacturer field rather than the name.
Counting live microphones is inference, not a query. CoreAudio cannot tell you how many transmitters are linked; the receiver enumerates as a fixed 2-channel USB device whether one lapel is on or both. But a channel with no transmitter carries true digital zero, while a linked transmitter sitting silent still sends its own self-noise, around -60 dBFS and never exactly zero. That gap is the whole discriminator. Presence is believed on the first buffer so the UI lights up instantly; absence must persist for a confirmation window, so an RF dropout cannot make the count flicker.
Reading device names needs no microphone permission. The CoreAudio HAL will name
every attached device before the user has granted anything, so Lapel can show
"DJI MIC MINI — 2 of 2 microphones live" on first launch and only prompt when you
actually press record. A browser-based equivalent cannot do this: getUserMedia
hides device labels until permission is granted.
swift build # build the library and the probe
swift test # full suite, no audio hardware required
swift run lapel-probe # live readout against real hardwareThe app icon is drawn by App/Icon/render-icon.swift; the asset catalog is
regenerated from its 1024 px output with sips.
Red/green TDD throughout — tests are written first and confirmed failing before any implementation exists. Every commit message says what was red. Tests use swift-testing.
MIT