Second-Order Models for Hidden Chunk Models

Inhalt:
Chunks are sequences of feature vectors representing segments, which are parts of phones. Hidden chunk models (HCMs) differ from HMMs in the acoustic model of segments leading to significant better performance. Still the HCM investigated hitherto assume that adjacent feature vectors are statistical independent. In this paper a second order approximation of emission probabilities is investigated, where the statistical dependency between two adjacent feature vectors within a chunk is modeled. Improvements compared to first order approximation have been achieved.