ESLASR aims to improve the quality of automatic speech recognition using loosely-coupled HMMs with articulatory-acoustic features.
This project addresses a shortcoming of current automatic speech recognition (ASR) systems - the way they deal with the processes of casual or fast speech, such as heavy co-articulation, vowel reduction and segment deletion. The scope for conventional HMMs-of-phones systems to cope with these processes is inherently limited because they use phone units. Co-articulation must be dealt with by using context-dependent models - a weak model of co-articulation. Vowel reduction and segment deletion are dealt with in the phonemic lexicon by listing alternate pronunciations, which cannot account for partial vowel reductions or feature spreading. We propose a combination of articulatory-acoustic features as a powerful representation of speech, and loosely-coupled HMMs as a suitable model to group these features into larger syllable units.
Edinburgh-Stanford link (Scottish Enterprise)