Detection of non-stationarity in speech signals and its application to time-scaling

This paper describes an automatic method for the detection of non-stationarity in speech signals. It is based on three measures of non-stationarity using Line Spectrum Frequencies (LSFs), the derivative of RMS values, and a combination of these two features. The application of the proposed method to time-scaling of speech signals is also presented. Results from an informal listening test support its usefulness. Following these results, the method seems to be a powerful tool for the automatic control of time-scale factors based on the characteristics of the input speech signal. Listeners preferred our new method over applying a constant time-scale factor in 90% of all cases. Other possible applications of the proposed tool are also discussed.