Voice data mining for laryngeal pathology assessment

The aim of this study was to evaluate the usefulness of different methods of speech signal analysis in the detection of voice pathologies. Firstly, an initial vector was created consisting of 28 parameters extracted from time, frequency and cepstral domain describing the human voice signal based on the analysis of sustained vowels /a/, /i/ and /u/ all at high, low and normal pitch. Afterwards we used a linear feature extraction technique (principal component analysis), which enabled a reduction in the number of parameters and choose the most effective acoustic features describing the speech signal. We have also performed non-linear data transformation which was calculated using kernel principal components. The results of the presented methods for normal and pathological cases will be revealed and discussed in this paper. The initial and extracted feature vectors were classified using the k-means clustering and the random forest classifier. We found that reasonably good classification accuracies could be achieved by selecting appropriate features. We obtained accuracies of up to 100% for classification of healthy versus pathology voice using random forest classification for female and male recordings. These results may assist in the feature development of automated detection systems for diagnosis of patients with symptoms of pathological voice.