A new algorithm based on RLS for voiced/unvoiced segmentation
Weibo Xie, Yongchu Wang, Canhui Cai · 2008
A new voiced/unvoiced segmentation based on RLS (recursive Least Square) in the foundation of traditional short-time analysis (i.e. Zero-Energy decision), with the adaptive tracking capacity to the non-steady pronunciation signal, has been present. The algorithm can rapidly realize precise voiced/unvoiced segmentation, without parameter-adjustment by samples training. In voiced section and unvoiced each one is relatively steady, the non-stability of pronunciation mainly manifests in the intersection point of voiced section and unvoiced section. In voiced/unvoiced section and the intersection point, RLS will present a different prediction track capacity inevitably. The adaptability establishes on a single pronunciation sample, and decides voiced/unvoiced segmentation according to the different tracking capacity of RLS in voiced/unvoiced section and the intersection point. It is different from the methods based on threshold, the algorithm based on recognizing the extreme value point of RLS prediction error would avoid the drawback of various adaptive learning algorithms in generalization, which better adapt for various varying factors of different sampling rate, speaker, volume, background noise.