$F_{0}$ estimation of speech using $L_{1}$-norm Sparse LP

Abstract

Linear Prediction (LP) analysis is currently used in speech processing that is based on $L_{2}$ -norm optimization. Nowadays, big data and data mining are paid deep attention, as a result, sparse estimation is strongly focused. Sparse estimation-based LP analysis has been proposed that is based on $L_{0}$ -norm or $L_{1}$ -norm optimization. We have already proposed Time-Varying Complex AR (TV-CAR) speech analysis for an analytic signal. These are based on $L_{2}$ -norm optimization and we have evaluated the $L_{2}$ -norm optimization method on speech processing such as $F_{0}$ estimation of speech, speech enhancement, HMM speech recognition. Complex residual resulting from inverse filtering for analysis signal is applied for $F_{0}$ estimation since the residual provides fewer components of formant structure. In this paper, the $L_{1}$ -norm based LP method is evaluated on $F_{0}$ estimation using the IRAPT (Instantaneous Robust Algorithm for Pitch Tracking) method in which the LP residual is used.