On an improvement of F0 estimation based on ℓ2-norm regularized TV-CAR speech analysis using AR pre-filter

Abstract

We have already proposed ℓ2-norm regularized TV-CAR speech analysis for an analytic signal that can suppress rapid spectral changes in time-domain and frequency-domain. We have already evaluated the performance using F0 estimation for noise corrupted speech with additive white Gauss noise or Pink noise. The IRAPT algorithm implemented the F0 estimation for the estimated complex AR residual from an analytic signal. We have found that a bone-conducted (BC) pre-filter makes it possible to improve the performance since the BC filter can suppress the additional noise. The BC characteristics are that of a low pass filter; as a result, a first-order AR filter can simulate the BC filter. This paper introduces first-order and second-order AR filters as the pre-filter to improve the F0 estimation performance.