出版社：Information and Media Technologies Editorial Board
摘要：This paper presents a new technique for recognizing musical instruments in polyphonic music. Since conventional musical instrument recognition in polyphonic music is performed notewise, i.e., for each note, accurate estimation of the onset time and fundamental frequency (F0) of each note is required. However, these estimations are generally not easy in polyphonic music, and thus estimation errors severely deteriorated the recognition performance. Without these estimations, our technique calculates the temporal trajectory of instrument existence probabilities for every possible F0. The instrument existence probability is defined as the product of a nonspecific instrument existence probability calculated using the PreFEst and a conditional instrument existence probability calculated using hidden Markov models. The instrument existence probability is visualized as a spectrogram-like graphical representation called the instrogram and is applied to MPEG-7 annotation and instrumentation-similarity-based music information retrieval. Experimental results from both synthesized music and real performance recordings have shown that instrograms achieved MPEG-7 annotation (instrument identification) with a precision rate of 87.5% for synthesized music and 69.4% for real performances on average and that the instrumentation similarity measure reflected the actual instrumentation better than an MFCC-based measure.
