A KLD Based Method for Initial Set Selection in Active learning

Speech recognition systems are usually trained using tremendous transcribed samples, and training data preparation is intensively time-consuming and costly. Aiming at achieving better performance of acoustic model with less transcribed samples, active learning is used in acoustic model training. This learning scheme firstly selects and transcribes a small initial training set, then iteratively selects the most informative samples corresponding to a certain criterion from the unlabeled samples, then annotates them and adds the newly transcribed samples to the training set to update the acoustic model. Concerning that the initial set influences the performance and convergence rate of active learning a lot, we proposed a method for initial set selection based on Kullback-Leibler Divergence (KLD). Our experiments show that active learning using initial set selected by our proposed method can achieve better performance.