Recently, the deep neural networks (DNNs) based acoustic modeling methods have been successfully applied to many speech recognition tasks. This paper reports the work about applying DNNs for syllable based acoustic modeling in Chinese automatic speech recognition (ASR). Compared with initial/finals (IFs), syllable can implicitly model the intra-syllable variations in better accuracy. However, the context dependent syllable based modeling set holds too many units, bringing about heavy problems on modeling and decoding implementation. In this paper, a WFST decoding framework is applied. Moreover, the decision tree based state tying and DNNs based models are discussed for the acoustic model training. The experimental results show that compared with the traditional IFs based modeling method, the proposed syllable modeling method using DNNs is more robust for data sparsity problem, which indicates that it has the potential to obtain better performance for Chinese ASR.
展开▼
