Estimation of Pitch of Noisy Speech using AR Model Based Inverse Filtering

This paper presents a noise robust method to extract the glottal pulse like signal and thereby estimating the pitch of a signal for a very wide range of signal-to-noise ratios (SNR). AR model based inverse filtering and discrete cosine transform (DCT) based pre and post-filtering has been done with the help of normalized autocorrelation model to enhance glottal pulse information of a noisy speech. This is further boosted by taking a time domain threshold, which results glottal pulse like signal with a period very close to the pitch period. The performance of the proposed method is evaluated by testing the method compare to two recent methods using standard TIMIT database. Test results show that the proposed method can estimate pitch with better accuracy and can also detect pitch reasonably accurately at very low SNR.