Time-varying autoregressive modeling approach for speech segmentation

Speech is considered as a nonstationary signal since the parameters such as amplitude, frequency and phase vary with time. Traditional speech segmentation is done based on a fixed frame length. However, speech characteristics can change within the fixed length or can be similar to the adjacent frames. Thus, it would be of interest to vary the length of the segment to accommodate the changes in the speech characteristics. The developed segmentation algorithm is based on a time-varying autoregressive model and the segmentation rules are developed based on the instantaneous energy and frequency estimate.