Automatic pronunciation modelling for multiple non-native accents
Silke Goronzy, K. Eisele · 2004
This paper describes an automatic method for generating non-native pronunciations and its combination with speaker adaptation to solve the problem of a performance decrease if state-of-the-art speech recognisers are faced with non-native speech. Although being a data-driven approach it overcomes the problem of gathering accented speech data for deriving the non-native variants. It rather uses solely native speech of both languages $the language the speaker is speaking as well as his/her mother tongue. Our experiments showed that using the generated accented variants to enhance the standard pronunciation dictionary in combination with weighted MLLR speaker adaptation outperforms the baseline system by up to 20% and speaker adaptation alone by up to 3%. The approach was tested on English-accented German and on German-accented French and results were consistent throughout both languages. Since our approach relies on native speech data only, it can easily be extended to various accents for different languages.