Application of Auditory Filter-banks in Polyphonic Music Transcription

Abstract

In this paper we present a frame-level transcription system for polyphonic piano music by using nonnegative matrix factorization (NMF) technique based on Fourier spectrogram as the representation of the musical signal and enhanced by application of an auditory filter-bank based on a new cochlear frequency-position equation, which was developed solving a biomechanical cochlea model without the need of physiological or psychoacoustic experiments. It is important to mention that in our days in music transcription task, a set of auditory bank filters have been used and this paper is focused precisely in this search field. Evaluation over a set of polyphonic piano pieces is performed to show the proposed method in this paper gives an increment in the precision measure of the music transcription in comparison when it is not used with NMF technique.