Robust pitch determination using nonlinear state-space embedding

A robust method for detecting periodicity and measuring fundamental frequency in speech and other signals is proposed. The method is based on concepts originally developed for analyzing chaotic time-series. A signal segment is transformed into trajectory in m-dimensional state space by using embedding procedure. Close pairs of points on the trajectory with the distance between them less than prescribed neighborhood radius are found and their time separations are computed. A periodicity histogram for the distribution of computed time separations is characterized by distinct peaks corresponding to pitch period and its multiples for periodic regions, and by the absence of such peaks for aperiodic regions. The proposed method does not suffer from the limitations of other short-term pitch-estimation techniques. Updated information and demo software can be found at www.soundmathtech.com/pitch.