Analysis of Closing-To-Opening Phase Ratio in Top-To-Bottom Glottal Pulse Segmentation for Psychological Stress Detection

Abstract

This paper is focused on investigating the differences in glottal pulses estimated by two algorithms; Direct Inverse Filtering (DIF) and Iterative and Adaptive Inverse Filtering (IAIF) for normal and stressed speech. Individual glottal pulses are mined from recorded speech signal and then normalized in two dimensions. Each normalized pulse is divided into a closing and opening phase and further segmented into npercentage sectors in Top-To-Bottom (TTB) amplitude domain. Three parameters, the kurtosis, skewness and pulse area, as well as their Closing-To-Opening phase ratios, are analysed. Designed GMM classifier is trained on speakers from Czech ExamStress database a further applied on other part of ExamStress database and also for English database SUSAS to investigate the independency of presented approach on spoken language and speech signal quality. The results achieved by DIF indicate independency on language and records quality (contrary to methods using IAIF). The best npercentage sectors in the TTB segments can be seen between 5 % and 40 %. In this case, methods based on DIF reached a psychological stress recognition efficiency of 88.5 % in average. The average stress detection efficiency of methods based on IAIF approached 73.3 %.This paper is focused on investigating the differences in glottal pulses estimated by two algorithms; Direct Inverse Filtering (DIF) and Iterative and Adaptive Inverse Filtering (IAIF) for normal and stressed speech. Individual glottal pulses are mined from recorded speech signal and then normalized in two dimensions. Each normalized pulse is divided into a closing and opening phase and further segmented into npercentage sectors in Top-To-Bottom (TTB) amplitude domain. Three parameters, the kurtosis, skewness and pulse area, as well as their Closing-To-Opening phase ratios, are analysed. Designed GMM classifier is trained on speakers from Czech ExamStress database a further applied on other part of ExamStress database and also for English database SUSAS to investigate the independency of presented approach on spoken language and speech signal quality. The results achieved by DIF indicate independency on language and records quality (contrary to methods using IAIF). The best npercentage sectors in the TTB segments can be seen between 5 % and 40 %. In this case, methods based on DIF reached a psychological stress recognition efficiency of 88.5 % in average. The average stress detection efficiency of methods based on IAIF approached 73.3 %.