Detection of Sigmatism with the aid of Machine Learning

The book is devoted to automatic detection of sigmatism in adult speech of German speakers. It has two major purposes: (1) to find an optimal set of audio features providing distinction between normal and disordered speech; (2) to create a Machine Learning (ML) classification algorithm able to analyze extracted features and detect sigmatism at phone level. The features are selected according to the phonetic background of considered sounds.They include first three formants, root-mean-square (RMS) amplitude, spectral peaks, spectral centroid, spectral skewness, and first 12 mel-frequency cepstral coefficients (MFCCs). Three ML methods are considered for sigmatism detection: Support Vector Machine, Gaussian Process, and Neural Networks. The process of feature extraction as well as automatic classification are conducted via Python scripts. As a result, the model based on SVM with the RBF kernel showed the highest accuracy rate of 90.6 %.