Phone boundary refinement using ranking methods

The HMM/SVM-based two-stage framework has been widely used for automatic phone alignment. The two-stage method uses SVM classifiers to refine the hypothesized boundaries given by the HMM-based Viterbi forced alignment. However, there are two drawbacks in using the classification model for detecting the phone boundaries. First, the training data contains only information about the boundary and far away non-boundary signal characteristics. Second, the classification model suffers from the class-imbalanced training problem. To overcome these drawbacks, we propose using ranking methods to refine the hypothesized boundaries. We train multiple phone-transition-dependent rankers by using K-means-based and decision-tree-based clustering. Both Ranking SVM and RankBoost are evaluated. The results of experiments on the TIMIT corpus demonstrate that the proposed ranking method outperforms the classification method. The best accuracy achieved is 84.20% within a tolerance of 10 ms. The mean boundary distance is 6.66 ms.