To improve the performance of automatic mispronunciation detection

a discriminative feature compensation training algorithm is proposed.The method is to train a matrix projecting from posteriors of Gaussians to a normal size feature space

and then to add the projected features to traditional spectral features.The matrix is trained according to maximum

F

1-score criterion

which aims at maximizing the empirical mispronunciation detection

F

1-score on the annotated speech database.Mispronunciation detection experiments have shown the method is effective in increasing

F

1-score

precision and recall on both the training data and evaluation data.It is

also shown model parameter discriminative training on new features obtained further improvements over both model training and feature training.