HHT based long term feature extracting method for speech emotion classification

Speech Emotion Recognition has become more and more important in the development of human-machine interactions. Long term features for the entire sentence are usually obtained by using statistical functions on the low-level descriptors such as pitch or energy at a frame level. In this paper, unlike the former statistical method, we propose a set of new long term features, denoted by LTF-HMS, based on HHT method. Compared to those statistic features based on the FFT, we tested what an important role the LTF-HMS plays in the experiment and the Berlin Emotional Database is applied as our corpus. The experimental results demonstrate that 39.5% of misclassification samples in shorter sentences and 25% misclassification samples in longer sentences caused by using FFT are identified correctly after integrating the LTF-HMS features.