Recognition of isolated musical patterns using discrete observation Hidden Markov models

Recognition of pre-defined musical patterns is very useful to researchers in Musicology and Ethnomusicology. This paper presents a novel efficient method for recognizing isolated musical patterns, using discrete observation Hidden Markov Models. The first stage of our method is to extract a vector of frequencies from the musical pattern to be recognized. This is achieved by means of a combination of a moving window technique with a largest-Fourier-peak selection algorithm. Each extracted peak frequency is subsequently quantized to a symbol of a finite and discrete alphabet. The resulting sequence of quantized frequencies is given as input to a set of Hidden Markov Models (HMM). Each HMM has been trained to model a specific pre-defined musical pattern. The unknown musical pattern is assigned to the model which generates the highest recognition probability. We have applied our method for the recognition of isolated musical patterns in the context of Greek Traditional Music. The resulting recognition rate was higher than 95%. Greek Traditional Music has distinct features, which distinguish it from the western equal-tempered tradition. To our knowledge, this paper presents a first effort for musical pattern recognition in the context of Greek Traditional Music using discrete observation Hidden Markov Models. Previous work was based on Dynamic Time Warping [8].