Spectralite Real-Time Beat Tracker

An online tracker hears only what has already played and must commit to each beat and tempo as the music arrives. It cannot go back and revise. An offline tracker reads the whole file, past and future, before it answers. Online tracking is the only kind that works for a live set, where the lights must fire on a beat before it lands. Every system on this page runs online on the same live audio, making the comparison like for like.

We introduce Spectralite RTBT, a streaming transformer for beat and tempo tracking. A causal convolutional front end reads the log-mel spectrogram. Transformer layers with rotary position embeddings then attend over a bounded window of past audio under a block-causal mask. Two output heads predict the beats and the local tempo directly, frame by frame.

RTBT learns from electronic and club music, with training built around the conditions of a live set: a DJ changing the tempo under it, and the room it plays in, from a venue's acoustics and crowd to a laptop microphone. It is tuned to pick up a tempo change quickly and to hold steady while the tempo holds.

For live lighting, RTBT announces each beat before it lands, early enough for a lighting desk to fire on it. Every beat score on this page counts a beat only when it was announced 100 ms ahead. The tracker needs only a small fraction of one laptop CPU core, and that cost stays the same however long the set runs.

We train RTBT to respond quickly when a DJ rides the pitch fader or jumps to the next record. The tracker predicts the tempo directly as well as the beats. It picks up a new tempo fast and settles onto the new beat grid.

The charts below measure the share of 256 clips where the beats line up again after a 4 % tempo change, and the share of the time the BPM shown is right.

A stable tracker holds the BPM steady through a track. It keeps the octave put and keeps the beats going through a breakdown. Spectralite RTBT folds octave errors to the tempo it already holds, so a 174 BPM track stays at 174. Genuine tempo changes within the octave pass straight through.

Annotators marked 151 breakdowns by hand in 95 full electronic tracks. In each one the kick drops out and comes back at the drop, and the tempo stays the same. The chart shows the share of labelled beats with a beat within 70 ms, announced 100 ms ahead, bar by bar around the start of the break and around the drop. The grey line at 100 % is a beat on every labelled beat.

The tracker maintains high accuracy on electronic music. It also holds up when the audio passes through a venue's acoustics, the crowd, and a laptop microphone.

Lighting consoles and DJ software hear the same audio as Spectralite RTBT through an audio loopback as it plays.

Tempo followed within 5 s and within 20 s is the share of tempo-change clips whose reported tempo settles within 1 % of the new tempo that soon after the change. Tempo right is the share of tracks whose reported tempo is within 4 % of the real one. A reading at half or double the tempo counts as wrong.

The numbers describe the whole system the tools run: the model, its decoder and the beat lock together.

Lighting software sends each cue early to cover the audio, network and fixture delay. These figures count a beat when it was announced at least 100 ms before it happened: Spectralite RTBT's scheduled beats, and for the other systems the beats their latest reported tempo and beat position predict. Systems that report a tempo without beats have no bar here. Each system's constant beat offset is measured once on separate calibration tracks and removed up to 40 ms either way. Only an offset beyond 40 ms counts against it.

Back on the beat means eight beats in a row arrive within 35 ms of the real beats. BPM right means the reported tempo is within 4 % of the music's while it holds steady. Half or double the tempo counts as wrong. Beats on time is the F-measure of the announced beats against the real ones within 70 ms, from 0 to 1. For the tempo-follow and ramp measures, the tempo is tracked within 1 %. Settled means the reported tempo is within 1 % of the new tempo and stays there within 2 s of the fader stopping.