A novel pitch cycle detection algorithm for tele monitoring applications

Recent advancements in the field of artificial intelligence has resulted in development of natural language processing using speech signals recorded over communication channels. Typically these speech recordings are segmented over a fixed frame length, usually 25ms, and used for training and testing mathematical models. For applications like speaker recognition and verification, this traditional methodology results in inclusion of lot of unvoiced and silence segments. To eliminate these redundancies, in this paper we propose a novel pitch synchronous segmentation algorithm robust to noise content and variations in speech recording parameters like sampling frequency. With two datasets coming from a handheld device and noise proof studio quality recordings, performance of this algorithm has been tested. With a maximum accuracy of 96.5%, this algorithm provides encouraging results.