Using accent information in ASR models for Swedish

A common technique to cope with the large variability in the acoustic realisations of the phonetic classes in speech, is to partition the data according to a linguistically significant variable. In this work, accent dependent phonetic models were trained and used both as an analysis tool for pronunciation variation and in the attempt to improve ASR performance. The Idea Accent dependent training The database is partitioned into accent areas. Accent dependent phonetic models are trained independently. speech database phone1 phoneN accentj accent dependent training accent dependent model set Accent analysis ⋆ The model parameters obtained this way, represent the statistical variation of the acoustic features across accent areas. This information can be used for pronunciation variation analysis. phonei clustering visual analysis