Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


Accent Specific Modifications for Continuous Speech Recognition Based on a Sub-Word Lattice Approach

K. Edwards, Fergus R. McInnes, Mervyn A. Jack

Centre for Speech Technology Research, Edinburgh, UK

Speech research has to date been dominated by either General American or RP British English accents. This limits the applicability of continuous speech recognition systems based on subword modelling to a small percentage of the English speaking population and makes true speaker independence more elusive. The modular speech recognition system discussed here allows experimentation at different stages in the recognition process, and is well suited to the implementation of different accents and languages. The 'front end' processing which uses Hidden Markov Models allows information about the phonetic and phonological details of an accent or language to be incorporated with the minimum of adjustment to the whole system. The 'back end' uses a lexicon which contains word transcriptions in terms of the units modelled by the HMMs. This paper describes the adaptation involved in enrolling speakers of a Scottish accent into the system, which was previously based on the assumption of RP accent users.

Full Paper

Bibliographic reference.  Edwards, K. / McInnes, Fergus R. / Jack, Mervyn A. (1991): "Accent specific modifications for continuous speech recognition based on a sub-word lattice approach", In EUROSPEECH-1991, 1195-1198.