Computer identification of wind instruments using cepstral coefficients

Abstract

Earlier results on computer recognition of the oboe and saxophone [J. C. Brown, J. Acoust. Soc. Am. 101, 3167(A) (1997); ‘‘Musical instrument identification using cepstral coefficients as features’’ submitted to J. Acoust. Soc. Am.] have been extended to include the clarinet and the flute. Approximately 30 samples of duration 2–10 s of each of the four instrument types comprise the test set. The training set consists of longer segments of approximately 1 min duration. Eighteen mel-based cepstral coefficients were calculated for each of the sounds. The training data were summarized by a cluster analysis with Gaussian probability density functions formed from the mean and variance for each of the clusters. The probability of belonging to each of the four classes was then calculated for the test sounds, and a Bayes decision rule was invoked to assign them to one of the classes. Results indicate that this is a very promising method for automatic recognition of musical instruments.