Abstract
This paper addresses accent issues in large vocabulary continuous speech recognition. Cross-accent experiments show that the accent problem is very dominant in speech recognition. Analysis based on multivariate statistical tools (principal component analysis & independent component analysis) confirms that accent is one of the key factors in speaker variability. Considering different applications, we proposed two methods for accent adaptation. When a certain amount of adaptation data was available, pronunciation dictionary modeling was adopted to reduce recognition errors caused by pronunciation mistakes. When a large corpus was collected for each accent type, accent-dependent models were trained, & a Gaussian mixture model-based accent identification system was developed for model selection. We report experimental results for the two schemes & verify their efficiency in each situation. 13 Tables, 5 Figures, 23 References. Adapted from the source document