Selection of relevant features for audio classification tasks

Abstract

Advances in time-frequency distributions and spectral analysis techniques (i.e., for the estimationof amplitude and/or frequency modulations) allow a better representation of non-stationary signalslike speech, highlighting their fine structure and dynamics. Although such representationsare very useful for analysis purposes, they complicate the classification tasks due to the largenumber of parameters extracted from the signal (“curse of dimensionality”). For such tasks, asignificant dimensionality reduction is required.In this thesis, the problem of dimensionality reduction of these time/frequency-frequencyrepresentations is studied; selection criteria of the optimal parameters are suggested, based ontheir relevance to a given classification task. Relevance is defined based on mutual information.First, using tools from multilinear algebra, such as High Order SVD, the initial dimensions andthe noise components of the representation are reduced. Then, feature selection proceeds basedon maximum relevance criterion. It is shown that the suggested process is equivalent to themaximum dependency criterion for feature selection, without, however, the need of the multivariateprobability densities estimation.The feature selection approach suggested in the thesis is applied on a number of audio classificationtasks, including speech detection in broadcast news and voice pathology detection anddiscrimination from vowel recordings. The complementarity of the modulation spectral features tothe state-of-the-art Mel frequency cepstral coefficients is shown for the above classification tasks.A system for the automatic discrimination of pathological heart murmurs using a high resolutiontime-frequency analysis of the phonocardiogram (PCG) is also presented. The classification accuracyof the system is comparable to the diagnostic accuracy of experienced paedo-cardiologistson the same PCG dataset.