A Cross Gender and Cross Lingual Study on Acoustic Features for Stress Recognition in Speech

We present a systematic study of the acoustic features for emotional stress in university students across gender and language groups. We design a common questionnaire of stress-inducing and non-stressinducing questions in Chinese and English, and interviewed 25 native speakers of Mandarin and 31 native speakers of English, of both gender. We extract 560 acoustic features including as low-level descriptors and Teager energy operator (TEO). Our acoustic feature-based classifier recognizes stress in the subjects’ speech with 81.28% accuracy on average within the same gender and language group, largely outperforming human perception tests which showed only 39.27%. Moreover, we show for the first time that whereas the emotion detection accuracy decreases by 28.18% across gender, our system maintains the same performance across Mandarin and English. Feature ranking experiments show that the most important stress features are TEO and MFCCs, rather than pitch. This explains the relative languageindependence of our model, even though Mandarin is a tonal language. TEO features are also founded to be insensitive to gender difference.