Applications of soft computing for musical instrument classification

Abstract

In this paper, a method for pitch independent musical instrument recognition using artificial neural networks is presented. Spectral features including FFT coefficients, harmonic envelopes and cepstral coefficients are used to represent the musical instrument sounds for classification. The effectiveness of these features are compared by testing the performance of ANNs trained with each feature. Multi-layer perceptrons are also compared with Time-delay neural networks. The testing and training sets both consist of fifteen note samples per musical instrument within the chromatic scale from C3 to C6. Both sets consist of nine instruments from the string, brass and woodwind families. Best results were achieved with cepstrum coefficients with a classification accuracy of 88 percent using a time-delay neural network, which is on par with recent results using several different features.