Novel active learning sample evaluation method based on multi-level confusion networks

Active Learning (AL) is designed to aid the labor-intensive process of training acoustic model for speech recognition. In AL, only the most informative training samples are selected for manual annotation. Thus, how to evaluate the unlabeled samples is worth researching. In this paper, we propose a unified framework to generate confusion networks of multiple levels including character, syllable and phone, and present a novel active learning sample evaluation method for Chinese acoustic modeling, posterior probabilities obtained from multi-level confusion networks are respectively adopted to evaluate the unlabeled samples. Our experiments show that compared with the widely used sample evaluation method using word posterior probability obtained from word confusion network, our proposed method can achieve satisfying performances.