How BeatScope Detects Key and Tempo

By the BeatScope analysis-engine team · updated July 14, 2026 · ~5 min read

In short: we compute tempo from an onset envelope and its autocorrelation, key from a chromagram correlated against Krumhansl–Schmuckler profiles, and the beat grid with dynamic programming. Everything runs offline, on-device. Here's exactly how, and why we chose these methods.

When we built BeatScope, the goal sounded simple: read three numbers off any track — BPM, key and energy — in seconds and without the internet, and do it accurately enough for DJs to trust in a live mix. Behind that "simple" sits classic digital signal processing. Here's our pipeline, honestly, step by step.

How we detect tempo (BPM)

Tempo is periodicity in a track's energy flow. We don't "listen for the kick" directly — we look for regularity in the moments when something happens in the sound.

1. Spectral flux → onsets. We measure how much the spectrum changes between frames. Sharp energy jumps are onsets (kick, snare, synth attacks). From them we build an onset envelope.
2. Autocorrelation. We autocorrelate that envelope: the lag where the signal most resembles itself is the beat period.
3. Comb estimation of multiple periods. We check not only the base period but its multiples (½, 2×), so tempo isn't confused with its harmonics.
4. Octave folding + range prior. We fold candidates into a 60–180 BPM range and gently prefer values typical of dance music — which makes the classic "174 vs 87" error rare.
5. Median smoothing. The final value is median-smoothed so single spikes don't throw the estimate off.

For the beat grid itself (where beats sit on the waveform) we use a dynamic-programming approach in the spirit of Dan Ellis's beat tracker^[3]: it finds a placement of beats that both lands on onsets and keeps a steady interval — so the grid doesn't drift across the track.

How we detect musical key

We estimate key from how energy is distributed across the 12 pitch classes — a chromagram.

1. Chromagram. We fold the spectrum (~55–2000 Hz) into 12 pitch classes and accumulate with a leaky integrator of about 8 seconds memory — giving an averaged "note profile" of the track.
2. Correlation with Krumhansl–Schmuckler profiles. We compare that profile with 24 reference key profiles (12 major + 12 minor). The best match is the key.
3. Mapping to Camelot. We return the result both in notes and in Camelot code — so harmonic mixing becomes arithmetic on the wheel.

This traces back to libKeyFinder by Ibrahim Sha'ath^[2] and to Carol Krumhansl's work on key perception^[1] — the same method studio key detectors use. We ported it into a pure on-device computation, with no server round-trip.

How we score energy

"Energy" is the sense of a track's drive. We estimate it from loudness (RMS) and spectral balance and map it to a simple scale, so you can order a set by rising intensity right in the library.

Accuracy and honest limits

On material with a clear rhythm section and stable tonality, the estimates are reliable. Where the method is objectively harder:

• Atonal or very dense harmony — the profile blurs, key is less certain.
• Mid-track modulation — we return the dominant key, not every shift.
• Weak percussion / rubato — few onsets, so BPM is less stable.
• Octave tempo errors — folding and the range prior suppress them, but on edge-case genres it's worth a manual double-check.

We deliberately don't oversell it: BeatScope is a fast, accurate assistant, and on the tricky cases the final call is always the DJ's ear.

Sources

1. Krumhansl C. L. Cognitive Foundations of Musical Pitch. Oxford University Press, 1990.
2. Sha'ath I. Estimation of Key in Digital Music Recordings. MSc thesis, Birkbeck College, University of London, 2011 (libKeyFinder).
3. Ellis D. P. W. Beat Tracking by Dynamic Programming. Journal of New Music Research, 36(1), 2007.