Automatic detection of speech inspiration using SVM and decision trees

Respiration is one of the most important subsystems in speech production. The detection and analysis of speech breathing patterns has numerous benefits in speech analysis, synthesis, recognition and in the development of voice biomarkers. Despite the close relation between speech and respiration, acoustic analysis of speech breathing is not well explored. In this paper, we propose a method for the automatic detection of inspiratory events from speech signals. This method is based on learning SVM and decision tree models using mel frequency cepstral coefficients (MFCC). We carried out an evaluation using a state-of-the-art database. The results showed, in particular, that fusion of the two models achieves an F1-score of 0.80, which significantly outperforms recently reported results.