Segmentation of Sindhi Speech using Formants

A speech segmentation method using formant frequencies is presented. The method uses speech samples of a major language of Indian sub-continent, Sindhi. It performs VCP (vowel-consonant-pause) segmentation and generates VCP strings for speech signals. The VCP strings and their formation may enable a recognizer to identify the speech on-the-fly, hence minimizing the system training and making the recognizer very efficient. The method applies velocity and acceleration parameters of rate-of-change dynamics on formants of speech to segment it into vowel, consonant, and pause parts. A test-bed software, to implement the proposed method and conduct all experiments, is also presented. Results show that the method is speaker as well as gender independent. Its segmentation performance is almost over 90% in most conditions and over 60% under some worst conditions. Long-term goal is to develop an efficient speaker-independent speech recognizer based on proposed method. A model of such a recognizer is also presented.