A spectrotemporal method based on Mutual Information (MI) is proposed for pitch estimation of voiced speech signals. We use MI as the similarity measure between voiced speech segments and their delayed version. In- stead of measuring linear dependencies, MI measures sta- tistical dependency, which suits the dynamic character- istic of speech signals. Besides, higher-order statistics are directly encoded in the MI while they are not usually taken into account in traditional correlation-based mea- sures. Through experiments on both synthetic signals and a real speech dataset, this new measure is proven to be ef- fective for pitch estimation.

Index Terms: pitch estimation, mutual information, higher-order statistics, periodicity, spectrotemporal