This paper examines the use of Mel-frequency Cepstral Coefficients in the classification of musical instruments. 2004 piano, violin and flute samples are analysed to get their coefficients. These coefficients are reduced using principal component analysis and used to train a multi-layered perceptron. The network is trained on the first 3, 4 and 5 principal components calculated from the envelope of the changes in the coefficients. This trained network is then used to classify novel input samples. By training and testing the network on a different number of coefficients, the optimum number of coefficients to include for identifying a musical instrument is determined. We conclude that using 4 principal components from the first 15 coefficients gives the most accurate classification results.