Deep Learning Based Language Modeling for Domain-Specific Speech Recognition

Abstract

Recurrent neural network (RNN) applied as language model of automatic speech recognition (ASR) can capture all previous history of word sequences and is theoretically superior to the N-gram language model. RNN has successfully applied in acoustic model of ASR and greatly improved the performance. In this paper, we combine RNN and N-gram language models and apply to a domain-specific speech recognition task. A rectification mechanism was introduced during RNN language model training procedure to prevent early stop and tried to deal with gradient vanishing issue. For Chinese word segmentation, a universal algorithm was designed for identifying proper name automatically, which improves word segmentation accuracy by 2.0–3.0%. Our experimental results show the new ASR system combing N-gram language model and domain-specific neural network language model can achieve lower word error rate than that with standard N-gram language model significantly.