Automatic Identification of Music Works Through Audio Matching

The availability of large music repositories poses challenging research problems, which are also related to the identification of different performances of music scores. This paper presents a methodology for music identification based on hidden Markov models. In particular, a statistical model of the possible performances of a given score is built from the recording of a single performance. To this end, the audio recording undergoes a segmentation process, followed by the extraction of the most relevant features of each segment. The model is built associating a state for each segment and by modeling its emissions according to the computed features. The approach has been tested with a collection of orchestral music, showing good results in the identification and tagging of acoustic performances.