Why guessing the BPM from a microphone will always drift.

What the microphone actually hears

It does not hear your track. It hears your track, plus the room, plus the crowd, plus the bar, plus whatever the monitors are throwing back, plus the reflection off the back wall arriving late. From that soup it has to find a pulse.

It usually manages. On a clean four-to-the-floor with a strong kick, tempo estimation is genuinely good. The trouble is that a set is not four minutes of clean kick.

Where it breaks, and it is never a convenient moment

• The breakdown. The kick goes away. The thing the estimator was locked onto disappears, and it starts hunting.
• The build. Snare rolls double and quadruple the apparent rate. Classic octave error: suddenly your visuals are running at double time.
• The drop. The exact instant that has to land perfectly is the instant right after the estimator was most confused.
• Your pitch fader. You ride the tempo, and a listener needs seconds to catch up. Seconds, on stage.
• The transition. Two tracks at once, two tempos, one answer expected.

Notice the pattern: estimation is worst precisely at breakdown, build and drop — the three moments that carry the whole show. It is reliable when nothing is happening and unsure when everything is.

Why "it looked fine in rehearsal"

At home, with a clean signal, no room and no crowd, listening works well enough that you ship it. The failure mode only shows up where you cannot debug it: loud room, live PA, someone talking into a mic, a monitor behind you.

What can actually be done about it

Two honest paths. The first is to keep listening but be humble about it: track the confidence, and refuse to change the tempo when you are not sure instead of chasing every ghost. A tool that says "I do not know" beats one that invents a number.

The second is to stop guessing. If the tempo already exists as a fact somewhere else in the booth, the right move is to read it rather than infer it — with your pitch already applied, and with no room, crowd or reflection in the way. That is the path we took, and it is why we could sit in front of real booth hardware and compare our number against theirs until the difference was zero.

It is not magic and it is not clever signal processing. It is just refusing to solve a hard problem when you can avoid having the problem.