Automatic silence/unvoiced/voiced classification of speech using a modified Teager energy feature
Alexandru Caruntu, Gavril Toderean, Alina Nica · 2005
Labeling of speech signals is a very important task, which can not miss in any of the early stages of developing a system based on a speech technology. Because of the large amount of time consumed when it is done by hand, there is a major need for algorithms that perform it automatically. This paper investigates a few methods of automatic classification of speech in silence/voiced/unvoiced (SUV) regions, using both time and frequency domain parameters. The features include zero-crossings, root mean square energy, but also a modified version of Teager energy, which proves to give the best results.