Design of a Musical Instrument Classifier System Based on Mel Scaled Cepstral Coefficient Supervectors and a Supervised Two-Layer Feedforward Neural Network

Abstract

A Content Based Audio Retrieval system is presented, which uses Mel ScaledF requency Cepstral Coefficients (MFCCs) as frequency based feature vectors to distinguish between different musical instruments via a Neural Network. Supervectors were formedfrom the concatenation of 14 dimensional feature vectors (12 MFCCs, 1 spectral centroid and1 frame energy feature) taken from 213 time slices, over 4 second samples. These feature sets were then usedto train a two-layer, feedforwardneural network, to segregate instrument samples into categories basedon their audio content. The motivation for this work is to create a system, which when presentedwith an unknown sample, can categorise it as belonging to one of a set of predefined classes. The system was tested on the recognition of instruments given 4 secondmonophonic phrases. An overall recognition accuracy of 40% was obtained, with drum and bass classes being the best categorised.