Discriminative Training For Automatic Mispronunciation Detection

In the field of speech recognition, under the guidance of discriminativetraining criterion to optimize the parameters of acoustic model has been widely used,the actual application to see the distinction between recognition performancestandards than production training standards have obvious improvement. Thesediscriminative criterion in order to reduce the expectations of the speech recognitionsystem training parameters error. Due to these characteristics of discriminativetraining, researchers began to study under the guidance of discriminative trainingcriterion optimizing the acoustic features by discriminative linear transformation, asto improve the recognition system. This method has been shown a good effectivenessin the speech recognition system.In pronunciation error detection, the system data should eventually with theevaluation of experts as consistent as possible, so the pronunciation error detectiontask in the performance evaluation index system is different from speech recognition.Due to these reasons these better effective discriminative training in speechrecognition is not have a ideal effectiveness in pronunciation error detection. F1valuefunction is a kind of discriminative training criterion that oriented pronunciation errordetection system. the recent experimental results show that compared with otherdiscriminative training criterion, the discriminative training F1was greatly increasedthe accuracy of pronunciation error detection system.Features directly affects the performance of error detection, thus starting fromthe feature extraction module to improve the pronunciation error detection system’sperformance is also a kind of method to improve the system. In this paper in order toimprove performance of pronunciation error detection system, try to used the idea thatoptimize the feature by linear transformation and discriminative training criterion inspeech recognition in the pronunciation error detection system. proposed adiscriminative feature training algorithm. The method is to train a matrix projecting from posteriors of Gaussians to a normal size feature space, and then to add theprojected features to traditional spectral features. The matrix is trained according tothe previously proposed maximum F1-score criterion, which aims at maximizing theempirical mispronunciation detection F1-score on the annotated non-native speechdatabase. Mispronunciation detection experiments have shown the method is effectivein increasing the F1-score, Precision and Recall on both the training data andevaluation data. It is also shown model space parameter discriminative training on thenew features obtained further improvements over both model-space training andfeature-space training.