Thesis (M.Eng.) - Asian Institute of Technology, 1995

A thesis submitted in partial fulfillment of the requirement for the degree of Master of Engineering.

In this study the talkspurt and silence pattern characteristics of conversational speech over telephone channel of three Asian languages : Chinese, Thai, Vietnamese have been investigated. The first goal is detection method. The proposed detection method has been tested with different threshold values to show its ability to eliminate noise that has the same nature of speech in this experiment. Although this method used to deal with this special noise, it can also be utilized for any other similar detection case with suitable minor changes of thresholds. Another aim of this study is to compare measured and modeled results of on-off speech pattern characteristics. Based on the detected data, probability density functions of talkspurt and silence duration without and with hangover or fill-in and statistical parameters are calculated to form a set of results named measured results. The silence duration pdf without hangover or fill-in modeled by using two weighted geometric function is used to derive silence duration pdfs with different hangover or fill-in, then, from these pdfs, statistical parameters are calculated to form a set of fitted results. Both measured and fitted results are compared and shown to be in reasonable agreement in small and mid-range of duration. The talkspurt duration pdfs with hangover or fill-in models do not give the high correlation coefficient . Therefore, a method using exponential function is proposed to model silence duration cdf and calculate statistical parameters to obtain a better agreement. The important of this work lies in the fact that the knowledge of conversational speech pattern would help in integration and interpolation of voice/data over bandwidth restricted systems which are being developed and deployed more and more in Asian countries.