Tracking Formant Trajectory of Continuous Chinese Whispered Speech with Hidden Dynamic Model Based on Dynamic Target Orientation

Abstract

Aimed at the characteristics of Chinese whispered speech formants, i.e., migrating to highfrequency, increased bandwidth, and increased spurious peaks and merged peaks, a method of tracking the formant trajectory of continuous Chinese whispered speech using the Hidden Dynamic Model (HDM) with dynamic target orientation was put forward in this study. The calculation proceeded as follows: firstly, the PIF-LPC algorithm was used to evaluate the formant parameters of whispered speech (PIF-LPC is an improved LPC algorithm. In PIF-LPC, pole interaction factors are used to correct the formant bandwidth of residual poles, to reduce the effect of pole intersection and to improve the accuracy of formant parameters); then, the extracted formant parameters as dynamic target orientation were introduced in HDM and compared with the actual observation results for realtime adjustment of the weight of dynamic target orientation; finally, HDM was solved through auxiliary particle filtering (APF), so as to realize the tracking of the formant trajectory of whispered speech. It was shown in the experimental results that the interferences of spurious peaks and merged peaks were avoided when the formant trajectory of continuous whispered speech was tracked by this method.