Musical timbre classification using FFT-Acoustic Descriptors and Machine Learning

Abstract

Musical timbre is a complex multidimensional attribute of auditory perception, which allows, in a first approximation, to discriminate between musical instruments when they have the same sound, intensity, and duration. Also, in some cases, there are sounds that appear to have very close timbral similarity, even when the instruments have different acoustic characteristics. This fact can make it difficult to classify musical instruments by timbres. We explore a 7dimensional abstract space, formed by the fundamental frequency and acoustic descriptors extracted from Fourier Transform in five musical instruments: Trumpet, violin, cello, transverse flute, and clarinet, of a monophonic audio record, from the Tinysol and Good-Sounds databases, corresponding to the fourth octave. This approach makes it possible to define a collection of points in timbral space uniquely and allows differentiating sounds played in ordinary style on any type of musical instrument. Through the geometric distance between musical sounds, we explore some Machine Learning techniques to establish categories of similarities between musical sounds, instruments, and family of musical instruments. It is concluded that the study of timbral similarity through geometric distances made it possible to find clustering between categories of musical timbre.