Continuous energy demodulation methods and application to speech analysis

Speech resonance signals appear to contain significant amplitude and frequency modulations. An efficient demodulation approach is based on energy operators. In this paper, we develop two new robust methods for energy-based speech demodulation and compare their performance on both test and actual speech signals. The first method uses smoothing splines for discrete-to-continuous signal approximation. The second (and best) method uses time-derivatives of Gabor filters. Further, we apply the best demodulation method to explore the statistical distribution of speech modulation features and study their properties regarding applications of speech classification and recognition. Finally, we present some preliminary recognition results and underline their improvements when compared to the corresponding MFCC results.