Contribution of Glottal Waveform in Speech Emotion: A Comparative Pairwise Investigation
This work provides insights into speech emotion recognition for affective computing applications, though it is incremental as it focuses on validating known components rather than introducing new methods.
The researchers investigated how much emotional information in speech comes from glottal waveforms by comparing emotion classification using full speech signals versus glottal-only signals in Mandarin Chinese, finding that glottal waveforms captured most emotional cues with up to 92.45% accuracy for distinguishing intense anger from moderate sadness.
In this work, we investigated the contribution of the glottal waveform in human vocal emotion expressing. Seven emotional states including moderate and intense versions of three emotional families as anger, joy, and sadness, plus a neutral state are considered, with speech samples in Mandarin Chinese. The glottal waveform extracted from speech samples of different emotion states are first analyzed in both time domain and frequency domain to discover their differences. Comparative emotion classifications are then taken out based on features extracted from original whole speech signal and only glottal wave signal. In experiments of generation of a performance-driven hierarchical classifier architecture, and pairwise classification on individual emotional states, the low difference between accuracies obtained from speech signal and glottal signal proved that a majority of emotional cues in speech could be conveyed through glottal waveform. The best distinguishable emotional pair by glottal waveform is intense anger against moderate sadness, with the accuracy of 92.45%. It is also concluded in this work that glottal waveform represent better valence cues than arousal cues of emotion.