ISCA Archive - Product code vector quantisation and hidden Markov modelling in isolated word recognition (original) (raw)

Product code vector quantisation and hidden Markov modelling in isolated word recognition

Borge Lindberg, Paul Dalsgaard

This paper presents a speaker independent isolated word recogniser, which combines the product codebook vector quantisation principle with the discrete hidden Markov modelling (HMM), so that each frame in the unknown test word (or training word) is described by two symbols, the linear predictive coding (LPC) shape and gain. The recogniser (both training and testing) has been evaluated on a 12 word vocabulary. The recognition results as well as the implementation requirements are discussed and compared with other approaches to speaker independent isolated word recognition.