Lapel

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 — dividing by the track count would halve the volume of a recording that never came close to clipping. The sum is taken at full weight, and a single gain applied across the whole mix only if its peak would exceed −1 dBFS.

The transcript is re-timed with the audio

Trim the mix and the transcript follows: turns outside the selection are dropped, and the rest are rebased so the two exported files still line up. A turn straddling a cut is clipped rather than dropped — someone mid-sentence at the boundary still said something inside the clip.

Mixing discards the separation, so the transcript carries it

Names you typed are used as written. An unnamed track becomes Speaker 1 rather than TX1 — 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.

Quieter is the whole answer

A lapel mic is a hand's breadth from its own speaker's mouth and a table's width from the other's. 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 doesn't hand the floor over, and a blip filter so a click on the wrong mic doesn't 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, genuine overlap is kept on both tracks, and it can be switched off for mics that could not hear each other.

Counting live microphones is inference, not a query

CoreAudio cannot tell you how many transmitters are linked — the receiver enumerates as a fixed two-channel USB device whether one lapel is on or both.

The signal can. A channel with no transmitter carries true digital zero. A linked transmitter sitting silent still sends its own self-noise, around −60 dBFS and never exactly zero. That gap is the entire discriminator.

The timing is deliberately asymmetric: presence is believed on the first buffer so the UI lights up the instant a lapel is switched on, while absence must persist for a confirmation window, so an RF dropout can't make the count flicker.

Reading device names needs no microphone permission

The CoreAudio HAL will name every attached device before you've granted anything, so Lapel shows “DJI MIC MINI — 2 of 2 microphones live” on first launch and only prompts when you actually press record.

A browser-based equivalent can't do this: getUserMedia hides device labels until permission is granted. It's one of several reasons this is Swift and not Electron.