Chroma and MFCC Based Pattern Recognition in Audio Files Utilizing Hidden Markov Models And Dynamic Programming
In this paper we present an algorithm to reveal the immanent musical structure within pieces of popular music. Our proposed model uses an estimate of the harmonic progression which is obtained by calculating beat-synchronous chroma vectors and letting a Hidden Markov Model (HMM) decide the most probable sequence of chords. In addition, MFCC vectors are computed to retrieve basic timbral information that can not be described by harmony. Subsequently, a dynamic programming algorithm is used to detect repetitive patterns in these feature sequences. Based on these patterns a second dynamic programming stage tries to find and link corresponding patterns to larger segments that reflect the musical structure.