Integrating Deep Neural Networks into Structured Classification Approach based on Weighted Finite-State Transducers

摘要

Recently, deep neural networks (DNNs) have been drawing the attention of speech researchers because of their capability for handling nonlinearity in speech feature vectors. On the other hand, speech recognition based on structured classification is also considered important since itsrealizes the direct classification of automatic speech recognition. For example, a structured classification method based on weighted finite-state transducers (WFSTs) introduces a linear classification term for each arc transition cost in a decoding network to capture contextual information from decoder states. In this paper, we focus on the integration of a WFST-based structured classifier and DNNs. Since these two approaches attempt to improve the representation of features and labels, respectively, the combination of these models would be efficient because of their complementarity. In the proposed method, DNNs are used to extract discriminative features, and then the features are classified by using WFST-based structured classifiers. The proposed method is evaluated by using TIMIT continuous phoneme recognition tasks. We confirmed that combining structured classification leads to stable performance improvements even from the welloptimized deep neural network acoustic models.