Recognition of the continuous-speech Russian phrases using their voiceless fragments

The author's a priori speech segmentation is applied for recorded signal as the basis to single out its voiceless fragments with increased reliability. We propose a continuous-speech recognition method by words partial lists generation from our full vocabulary using the number of their voiceless fragments. The appropriate parts of the signal are recognized in these dictionaries by evaluation of DTW-distance to the word patterns which are synthesized from diphone patterns. Experiments were performed on vocabularies with several hundreds of nouns and adjectives of highest frequency, recorded by the Russian language frequency dictionary. In these experiments, the correct recognition was over 90%.