Development of neural network-based speech/non-speech discrimination algorithm for audio files

Abstract

For research in speech processing and analysis of audio content in general, extensive data sets are required. Creating such a dataset manually turns out to be time and labour consuming, but it is a task very suitable for automation through machine learning. This dissertation describes the development of an algorithm for audio content analysis, discriminating between speech and music audio classes. To detect the different classes, the audio signal needs to be expressed in terms of its statistical properties based on spectral and temporal features and statistical values of these features are used to differentiate between the audio classes. The suitability of various low-level audio features was evaluated to determine suitability and efficiency in discriminating between different audio classes. To gain a better understanding of the feature set, exploratory dimensionality reduction analysis was performed with Principal Component Analysis (PCA). The mean accuracy for the SVM classifier was used to rank combinations of features. A number of feature selection techniques were employed to reduce the feature space and increase the mean accuracy, these included Univariate feature selection, Random Forest Regression, forward and backward Sequential Feature Selection and the highest loading factors of the PCA components. An optimal subset of features was selected and used in evaluating a Neural Network-based classifier model and validated against a Support Vector Machine model. It was demonstrated that using this novel method of selecting the optimal combination of audio features, a 50% reduction in dimensionality and higher mean accuracy (99.95%) was achieved and proved to be well suited as a tool for extracting and compiling speech data sets. The algorithm developed in this study is pertinent and applicable to the requirement of the initial motivation for the study, to efficiently create datasets for the study of language identification and recognition for African and other indigenous languages and beyond that, for analysing audio content in general. The reduction in the dimensionality of the feature space and consequently reduction in computational load should capacitate a real-time audio content analysis tool, implemented on low power IoT devices.