On an Improved F0 Estimation Based on ℓ2-Norm Regularized TV-CAR Speech Analysis

Spectral estimation performance determines that of speech processing. Linear Prediction (LP) is the most successful speech analysis method commonly introduced worldwide of a smartphone, LINE, Skype to realize CELP speech coding to extract the spectral features with a small amount of computation and fewer parameters. Besides the CELP coding, the LP performs better on the
$F_{0}$
, estimation since the LP residual contains fewer formant structures. We have already proposed a time-varying complex AR (TV-CAR) speech analysis for an analytic signal that estimates the time-varying complex AR parameters from a speech signal. We recently proposed the TV-CAR analysis based on ℓ
2
-norm regularized LP and evaluated the effectiveness of the performance on the
$F_{0}$
estimation using the IRAPT algorithm. On the other hand, bone-conducted (BC) speech is robust against additive noise since it provides a stable harmonic structure in low frequencies and cannot be easily affected by the noise. In this paper, we introduce a pre-filter that simulates BC speech to improve the performance of the
$F_{0}$
, estimation. The experimental results show that the BC filter improves the performance for the high level of noise corrupted female speech.