Estimation of syntactic structure using prosodic parameters and its application to speech recognition

Prosodic information is very useful for the estimation of syntactic structure. So Phrase-level segmentation of continuous speech using pitch and power information is discussed to utilize for the word spotting. First, a rule based approach using pitch and power pattern is carried out. Then, a statistical approach is carried out, where the vector sequence of pitch and power are matrix quantized and conditional probabilities of the codes and the labels are calculated. The experimental result of the rule-based approach shows that the boundary detection accuracy is 72.8% in case that the adjacent phrases have loose relation, and 43.4% in case of tight relation. As for the statistical method, the boundary detection accuracy is 63.0%.