Musical beat recognition using a MLP-HMM hybrid classifier

This paper describes a system for detecting the musical beats in popular music. The beat recognition system is designed to follow how humans detect drum onsets and perceive beats. Mel-frequency cepstral coefficients (MFCCs) are extracted from the sample music. These MFCCs are used by a hybrid multi layer perceptron-hidden Markov model classifier to extract bass and snare drum onset locations from the musical input. From these locations, multiple agents with tempo hypotheses are created. Agents increase in score as they predict beats correctly. The highest scoring agent's beats are considered the correct beats. These output beats are compared to a hand-transcribed ground truth made using visual and aural inspection of the input music's waveform. A beat recognition accuracy of 74.56% was obtained using a test set consisting of 30-second song samples with drums.