Musical Instrument Identication with Feature Selection Using Evolutionary Methods
Bernadette Loughran · 2009
Musical Instrument Identification with Feature Selection Using Evolutionary Methods Roiśin Loughran Musical instruments may be identified using machine learning methods, but it is not clear which aspects of the sound or features are best used in such methods. Classification experiments using Principal Component Analysis (PCA) and Multi-Layered Perceptrons (MLP) in this thesis find that the addition of extra features may not necessarily be beneficial — optimisation of the features is required. This optimisation is implemented using Evolutionary Computation methods as they have yet to be extensively applied in musical sound analysis. A Genetic Algorithm (GA) with a new instrument-clustering fitness function based on PCA is applied to optimise a set of 95 features for classification with an MLP. With this method, the number of features used to classify an instrument is reduced from 95 to as low as 22 with a classification accuracy reduction of less than 0.3%. This method is tested against another evolutionary method that has not yet been applied to instrument identification — Genetic Programming (GP). GP is used to evolve a classifier program that can identify unseen samples with an accuracy of 94.3% using just 14 of the 95 original features. Though not as high as the MLP or the GA-MLP, it is found that GP is more consistent with its choice of features, offering a possible insight into timbre and the nature of sound recognition. In both EC methods it is found that the first principal component of the envelope of the centroid, a new measure of this feature, is the most important among all 95 features. It is also seen that each classification method performs significantly better when tested with a general set of samples, than with a oneoctave sample set common to each instrument. The classifiers are compared to a set of human listening tests on particularly troublesome samples. It is seen that although the GA and GP are accurate at identifying general unseen samples, the human ear performs significantly better than both methods at identifying these difficult samples.