Vocal Emotion Recognition with Log-Gabor Filters

Vocal emotion recognition aims to identify the emotional states of speakers by analyzing their speech signal. This paper builds on the work of Ezzat, Bouvrie and Poggio by performing a spectro-temporal analysis of affective vocalizations by decomposing the associated spectrogram with 2D Gabor filters. Based on the previous studies of the emotion expression in voices and the turn out display in spectrogram, we assumed that each vocal emotion has a unique spectro-temporal signature in terms of orientated energy bands which can be detected by properly tuned Gabor filters. We compared the emotion-recognition performances of tuned log-Gabor filters with standard acoustic features. The experimental results show that applying pairs of log-Gabor filters to extract features from the spectrogram yields a performance that matches the performance of an approach based on traditional acoustic features. Their combined emotion recognition performance outperforms state-of-the-art vocal emotion recognition algorithms. This leads us to conclude that tuned log-Gabor filters support the automatic recognition of emotions from speech and may be beneficial to other speech-related tasks.