Speech compaction using vector quantisation and hidden Markov models

Abstract

We present techniques for the time compaction of speech using vector quantisation and hidden Markov modelling. These aim to retain the most perceptually important information present in the speech signal, while discarding redundant information. The methods are compared with the conventional technique using synchronised overlap-add (SOLA) compaction, and with a recently proposed hierarchical temporal decomposition (HTD) based method. Using mean opinion score testing, they are found to give a better quality output than the SOLA method, and similar quality to the HTD.