A confusion network based confidence measure for active learning in speech recognition

Speech recognition systems are usually trained using tremendous transcribed utterances, and training data preparation is intensively time-consuming and costly. Aiming at reducing the number of training examples to be labeled, active learning is used in acoustic modeling of speech recognition, this learning scheme iteratively inspects the unlabeled samples, selects the most informative samples corresponding to a certain criterion, then annotates them, and adds the newly transcribed samples to the training set to update the acoustic model. Concerning about the importance of the criterion to select the most informative samples, we proposed a confidence measure computed by confusion network, and used this measure as the criterion for sample selection to improve the efficiency of active learning in acoustic modeling. Our experiments show that active learning, which adopts the proposed confidence measure, can achieve 31% maximum reduction of labeled data compared with random selection method.