电话交谈语音识别中基于LSTM-DNN语言模型的重评估方法研究

Revaluation based on LSTM- DNN language model in telephone conversation speech recognition

近年来,神经网络语言模型的研究越来越受到学术界的广泛关注。基于长短期记忆(long short-term memory,LSTM)结构的深度神经网络(LSTM-deep neural network,LSTM-DNN)语言模型成为当前的研究热点。在电话交谈语音识别系统中,语料本身具有一定的上下文相关性,而传统的语言模型对历史信息记忆能力有限,无法充分学习语料的相关性。针对这一问题,基于LSTM-DNN语言模型在充分学习电话交谈语料相关性的基础上,将其应用于语音识别系统的重评估过程,并将这一方法与基于高元语言模型、前向神经网络(feed forward neural network,FFNN)以及递归神经网络(recurrent neural network,RNN)语言模型的重评估方法进行对比。实验结果表明,LSTMDNN语言模型在重评估方法中具有最优性能,与一遍解码结果相比,在中文测试集上字错误率平均下降4.1%。

In recent years,the research on the neural network language model has received more and more attention from the academic circles. At present,the neural network language model based on LSTM structure has become a research hotspot. In the speech recognition system,the corpus itself has certain relevance. But the traditional language models have limited memory capacity,and they cannot fully learn the relevance of the corpus. To solve this problem,a novel LSTM-DNN language model is applied to the revaluation of speech recognition,which fully exploits the correlation on a telephone conversation corpus. It is further compared with existing revaluation methods based on language models such as high order language model,feed forward neural network（ FFNN） language model and recurrent neural network（ RNN） language model. The experimental results show that the performance of LSTM-DNN language model is optimal. Compared to the first pass of decoding,the relative decline in average word error rate is 4. 1% in the Chinese test sets.