An accurate timbre model for musical instruments and its application to classification

Abstract. A compact, general and accurate model of the timbral char-acteristics of musical instruments can be used as a source of a priori knowledge for music content analysis applications such as transcription and instrument classification, as well as for source separation. We de-velop a timbre model based on the spectral envelope that meets these requirements and relies on additive analysis, Principal Component Anal-ysis and database training. We put special emphasis on the issue of fre-quency misalignment when training an instrument model with notes of different pitches, and show that a spectral representation involving fre-quency interpolation results in an improved model. Finally, we show the performance of the developed model when applied to musical instrument classification.