Comparison of LPC estimation algorithms from noisy speech

Abstract

It is well known that LPC parameters are severely distorted when the speech waveform is corrupted by noise. This paper compares the performance of different LPC estimation algorithms from noisy speech. Direct estimation of LPC parameters from noisy speech as well as the use of preprocessor speech enhancement algorithms are used. The direct estimation algorithms considered are: the autocorrelation subtraction method, shifted Yule‐Walker method, and an instrumental variable method. The preprocessor speech enhancement algorithms considered are: adaptive filtering (AF), adaptive noise cancellation (ANC), and linear maximum a priori (LMAP) estimation. The algorithms are applied on speech corrupted by white noise. Their performance is compared through Monte‐Carlo simulation conducted on voiced, unvoiced, and v/uv mixture of frames. The criteria chosen for comparison are bias, standard deviation, and average euclidean distance. The above algorithms are also applied to a speaker recognition scheme based on orthog...