Robust linear prediction based on adaptive lattice filter

A robust speech-analysis method based on an adaptive weighted least-squares lattice algorithm is introduced to improve the performance of conventional frame-based linear predictive coding analysis. The likelihood variable, which can be considered as a statistic measure of nonGaussian components, is used to de-weight the excitation interval. The likelihood variable can also be considered as a useful statistic for detecting outliers. Short analysis windows, with optimal frame position placement determined according to local minimum of the likelihood variable and residual, are adopted to mimic glottis-closure analysis. Experiments on both synthetic speech and real speech are included to show that the algorithm offers improved formant estimation accuracy in comparison to the frame-based method