Contributions to automatic multiple F0 detection in polyphonic music signals

Abstract

Multiple fundamental frequency estimation, or multi-pitch estimation (MPE), is a key problem in automatic music transcription (AMT) and many other related audio processing tasks. Applications of AMT are numerous, ranging from musical genre classification to automatic piano tutoring, and these form a significant part of musical information retrieval tasks. Current AMT systems still perform considerably below human experts, and there is a consensus that the development of an automated system for full transcription of polyphonic music regardless of its complexity is still an open problem. The goal of this work is to propose contributions for the automatic detection of multiple fundamental frequencies in polyphonic music signals. A reference MPE method is chosen to be studied and implemented, and a modification is proposed to improve the performance of the system. Lastly, three refinement strategies are proposed to be incorporated into the modified method, in order to increase the quality of the results. Experimental tests reveal that such refinements improve the overall performance of the system, even if each one performs differently according to signal characteristics.