The generation of regional pronunciations of English for speech synthesis

Welsh and Northern English), and two American ones (New York and South Carolina, to represent Eastern and Southern American); regional features were based primarily on the descriptions in [1], with native-speaker input where possible. The regional accents are abbreviated in this paper as: Br(Sc) = Edinburgh; Br(W) = Cardiff; Br(N) = Leeds; Am(E) = New York; and Am(S) = South Carolina. For the standard accents, Br(RP) = RP, and Am(Gen) = General American. Most speech synthesisers and recognisers for English currently use pronunciation lexicons in standard British or American accents, but as use of speech technology grows there will be more demand for the incorporation of regional accents. This paper describes the use of rules to transform existing lexicons of standard British and American pronunciations to a set of regional British and American accents. The paper briefly discusses some features describes of the regional accents in the project, and the framework used for generating pronunciations. Certain theoretical and practical problems are highlighted; for some of these, solutions are suggested, but it is shown that some difficulties cannot be resolved by automatic rules. However, although the method described cannot produce phonetic transcriptions with 100% accuracy, it is more accurate than using letter-to-sound rules, and faster than producing transcriptions by hand. The accents generated represent fairly educated regional speech, though some optional rules were included which produce broader accents. The division between 'obligatory' and 'optional' rules is somewhat artificial, as there may be speakers from the region who have a noticeably local accent but do not use all of the 'obligatory' rules as their speech is somewhat closer to the standard accent. However, it enables us to produce pronunciation lexicons which represent the main features of the regional accents, while allowing some freedom of variation.