2004 IEEE Workshop on Machine Learning for Signal Processing REPRESENTING TIMBRE DYNAMICS OF A M US IC AL I N ST RUMEN T: C 0 M PAR IS 0 N BETWEEN GA AND PCA Hugo B. de Paula, H. C. Yehia, J. A. Vasconcelos and M. A. Loureiro CEFALA-Center for Research on Speech, Acoustics, Language and Music AV Pres. .4nt6nio Carlos, 6627 - 31270-010 - Belo Horizonte-MG - Brazil E-mail: hugohp, hani&efala.org; jvasconcelos, mauricioloureiro@ufmg.br Web : uww.cefala.org Phone: +55 (31) 3499-4937 Fax: +55 (31) 3499-5480 Abstract. The Time-varying spectral characteristics of clarinet sounds were used to build maps with the’purpose of representing the great variety of sounds a musical instrument may produce. A database of notes performed in several intensities and covering all the extension of the clarinet was used. This work compares two kinds of orthogonal bases used to create a sound map from time- varying curves of sound partials. The first mapping was derived by Principal Component Analysis and the second was created us- ing Multiple Wavetable Synthesis (MWS) and Genetic Algorithms. These bases defined spectral subspaces capable of representing and grouping all tested sounds, which were validated by auditory tests. Sub-spaces involving larger groups of notes were used to compare the sounds according to the distance metrics of the representation. INTRODUCTION Representation of a musical instrument,involves the est.imat.ion of the physical parameters that contribute to t.he perception of pitch, int.ensit,y and tinihre of all sounds the instrument is capable of producing. Due to its inherently multidimensional nature, timbre cannot be easily scaled. It is perceived hy means of the interaction of a \.aricty of sbatic and dynamic properties of sound grouped into a complex set of auditory att,ributes. The identificat,ion of the contribution of each one of these competitive factors has been t,he main subject of psychoacoustics research on timbre percept,ion. The introduction of the notion of “similarity rate“ of hearing judgment responses t,ogether wiih bIult.idimensiona1 Scaling (MDS) techniques allowed the reduction of t,his multidimensionality and made it. possible to investigate t,he complex structure of this attribute [ll]. In one of the most classic stud- ies on musical timbre, Grey 141 measured subjective judgment of similarity 0-7803-8608-6/04/$20.00 02004 IEEE 381