Perceptual Time Varying Linear Prediction model for speech applications

A new perceptual time varying model for non-stationary analysis of speech signals is presented. Some researches have already shown that the time varying linear prediction coding (TVLPC) model that was applied to speech signals increases the recognition performance of automatic speech recognition (ASR) systems. This improvement has been achieved due to the incorporation of the speech dynamics information in the model. Another work, perceptual linear prediction (PLP) analysis of speech, has shown that a modified estimation of the auto correlation function (ACF) of stationary speech frame yields major improvement to the recognition rate. The presented model, perceptual time varying linear prediction (PTVLP) analysis of speech, adopts the perceptual concepts, of how to estimate the ACF, into the TVLPC model. This research shows that the proposed PTVLP model is more accurate, robust to noise and achieves better recognition rates than PLP and TVLPC over wide SNR range.