Classification of audio signals using spectrogram surfaces and inherent distortion measures

Abstract

Representation of one-dimensional (1D) signals as surfaces and higher dimensional manifolds reveals geometric structures that can enhance assessment of signal similarity and classification of large sets of signals. We therefore represent 1D signals as surface objects embedded in higher dimensional Euclidean or other spaces. Specifically, since we are concerned with audio signals, the spectrogram is utilized for the representation of the 1D signals by surfaces, but a handful of other representations in combined spaces such as wavelets can be utilized for this purpose. A novel class of geometric features is then extracted by parameterizing the surfaces, and by utilizing distortion measures that are defined with reference to them. This yields a set of highly descriptive features that are instrumental in our approach to feature engineering, analysis and classification of audio or other 1D signals. Two examples of audio signals were selected to illustrate applications and the capabilities of the new approach: Lung sounds were chosen in view of the interest nowadays in respiratory pathologies caused by the Corona virus and environmental problems; Accent detection was selected as an example of a challenging speech problem. The proposed approach outperformed baseline models under all measured metrics. Our novel approach to 1D signal representation and processing can be further extended considering higher dimensional distortion measures.