Robust pronunciation evaluation in adverse environments

Pronunciation evaluation systems used by many people in the same place at one time need to evaluate the pronunciation robustly. In order to deal with the robust problem, this paper first applies multi-training plus adaptation for acoustic models refinement as in robust speech recognition and then introduces a nonlinear mapping method using CDF-matching for the evaluation feature normalization. Experimental results indicate that multi-training and adaptation can improve the performance. At the same time, the nonlinear feature mapping method still yields a performance improvement.