Speaker Independent Connected Digit Recognition Using Diphone-Like Units with HMM Techniques

A subword unit approach to speaker independent connected digit recognition using Hidden Markov Model (HMM) techniques has been investigated. The subword units chosen for this study are hybrid single-phone/diphone units proposed for both template model and HMM approaches to speech recognition. This investigation is an important first step in the process of applying the approach to large vocabulary continuous speech recognition tasks. Although there are many possible choices for subword units for recognition purposes, it is felt that a mixed composition of single phone and diphone speech sounds provides a good compromise between maintaining a manageable inventory of speech sound units and providing more stable and representative models of speech sounds than can be achieved by using only single phones.