Accent features and idiodictionaries: on improving accuracy for accented speakers in ASR

One of the most widespread approaches to dealing with the problem of accent variation in ASR has been to choose the most appropriate pronunciation dictionary for the speaker from a predefined set of dictionaries. This approach is weak in two ways: firstly that accent types are more numerous and more variable than can be captured in a few dictionaries, even if the knowledge were available to create them and secondly, accents vary in the composition and phonotactics of the phone inventory not just in which phones are used in which word.;In this work, we identify not the speaker's accent, but accent features which allow us to predict by rule their likely pronunciation of all words in the dictionary. Any given speaker is associated with a set of accent features, but it is not a requirement that those features constitute a known accent. We show that by building a pronunciation dictionary for an individual, an idiodictionary, recognition accuracy can be improved over a system using standard accent dictionaries.;The idiodictionary approach could be further enhanced by extending the set of phone models to improve the modelling of phone inventory and variation across accents. However an extended phoneme set is difficult to build since it requires specially-labelled training material, where the labelling is sensitive to the speaker's accent. An alternative is to borrow phone models of a suitable quality from other languages. In this work, we show that this phonetic fusion of languages can improve the recognition accuracy of the speech of an unknown accent.;This work has practical application in the construction of speech recognition systems that adapt to speakers' accents. Since it demonstrates the advantages of treating speakers as individuals rather than just as members of a group, the work also has potential implications for how accents are studied in phonetic research generally