Beat detection is a multifaceted problem in music information retrieval (MIR) that involves automatically extracting rhythmic structure from audio signals—a task that human listeners perform intuitively but that proves surprisingly challenging for algorithms. The fundamental difficulty lies in the fact that rhythm and tempo are perceptual constructs that emerge from complex interactions between acoustic features, and no single signal property directly encodes the beat.

The most common approach to automatic beat detection begins with onset detection—identifying the precise moments where new musical events begin. Onsets correspond to the attacks of notes, drum hits, and other transient events that create the rhythmic backbone of music. Onset detection algorithms typically compute a function that measures spectral change over time, called an onset detection function or novelty function. This function peaks whenever the audio spectrum changes significantly—when a new drum hit arrives, a chord changes, or a note begins. Common methods include computing the spectral flux (the sum of positive differences between successive spectral frames), the complex domain deviation, or the high-frequency content function that emphasizes the sharp transients characteristic of percussive sounds.

Once onsets are detected, the next step is periodicity estimation—finding the regular temporal interval that best explains the pattern of onsets. The autocorrelation function is a classical tool for this purpose: it measures the similarity of the onset detection function with time-shifted copies of itself, revealing periodicities as peaks in the autocorrelation output. The location of the dominant peak corresponds to the most prominent periodic interval, which is then converted to BPM. Alternative approaches include tempo histograms, which accumulate inter-onset intervals into bins and identify the most common interval, and comb filter banks, which test the onset function against templates of regularly spaced impulses at different tempos.

A persistent challenge in BPM detection is tempo ambiguity—the mathematical reality that any tempo has valid multiples and subdivisions. A track at 120 BPM contains equally valid periodicities at 60 BPM (half-time) and 240 BPM (double-time). Sophisticated algorithms address this through musical heuristics, tempo range priors (most popular music falls between 80-160 BPM), and multi-level analysis that considers multiple metrical levels simultaneously. The confidence score reported alongside BPM estimates reflects how strongly the detected periodicity stands out against competing candidates, with high-confidence results typically indicating music with a clear, consistent rhythmic pattern and lower confidence suggesting complex, irregular, or tempo-varying musical content.