Speech Segmentation Algorithm Based on an Analysis of the Normalized Power Spectral Density

This article demonstrates a new approach to speaker independent phoneme detection. The core of the algorithm is to measure the distance between normalized power spectral densities in adjacent, short-time segments and verify it based on velocity of changes of values of short-time signal energy analysis. The results of experiment analysis indicate that proposed algorithm allows revealing a phoneme structure of pronounced speech with high probability. The advantages of this algorithm are absence of any prior information on a signal or model of phonemes and speakers that allows the algorithm to be speaker independent and have a low computation complexity.