Classifying females' stressed and neutral voices using acoustic-phonetic analysis of vowels: an exploratory investigation with emergency calls

Abstract

In the present exploratory study, we investigated acoustic–phonetic measures of spoken vowels for detection of female speech under conditions of stress. Eight authentic recorded calls to emergency services received from eight Finnish adult female speakers were chosen for the analysis. Based on the purpose of the call, the recordings were divided into two groups: the stressed group and the neutral group. We chose f0, H1–H2 and centre of gravity as acoustic–phonetic predictors for our final classification models; In previous studies, high f0 has been associated with a stressed voice, but H1–H2 and centre of gravity have not previously been related to speech under stress. On the other hand, H1–H2 has been used to detect non-modal voice qualities, such as a creaky or breathy voice, and similar voice qualities have been observed in stressed speech. Furthermore, indications exist that in speech under stress, acoustic energy is concentrated in higher frequencies, which consequently increases the centre of gravity. We tested stress detection accuracy with three statistical classifiers: LDA, logistic regression and decision tree. Our results indicated that all the models performed better when they were trained using only the vowel /i/ rather than training them with all Finnish vowels. The use of our best performing model, a logistic regression model based on /i/, yielded 88% correct classification, whereas a logistic regression model trained with all vowels achieved an accuracy of only 81%. We conclude that the results indicate a good stress classification accuracy, although further research with more extensive data is required.