基于RNN-RBM语言模型的语音识别研究

发布时间：2018-03-19 08:47

本文选题：语音识别　切入点：语言模型　出处：《计算机研究与发展》2014年09期 　论文类型：期刊论文

【摘要】：近年来深度学习兴起,其在语言模型领域有着不错的成效,如受限玻尔兹曼机(restricted Boltzmann machine,RBM)语言模型等.不同于N-gram语言模型,这些根植于神经网络的语言模型可以将词序列映射到连续空间来评估下一词出现的概率,以解决数据稀疏的问题.此外,也有学者使用递归神经网络来建构语言模型,期望由递归的方式充分利用所有上文信息来预测下一词,进而有效处理长距离语言约束.根据递归受限玻尔兹曼机神经网络(recurrent neural network-restricted Boltzmann machine,RNN-RBM)的基础来捕捉长距离信息;另外,也探讨了根据语言中语句的特性来动态地调整语言模型.实验结果显示,使用RNN-RBM语言模型对于大词汇连续语音识别的效能有相当程度的提升.
[Abstract]:In recent years, the rise of in-depth learning has a good effect in the field of language modeling, such as restricted Boltzmann machine restricted Boltzmann model, which is different from N-gram language model. These neural network-based language models can map word sequences to a continuous space to evaluate the probability of the next word appearing to solve the problem of sparse data. In addition, some scholars use recurrent neural networks to construct language models. It is expected that all the above information can be used recursively to predict the next term and thus deal with long distance language constraints effectively. The long distance information is captured according to the recurrent neural network-restricted Boltzmann machine RNN-RBM, which is based on the recurrent neural network-restricted Boltzmann machine neural network of recursive constrained Boltzmann machine. This paper also discusses how to dynamically adjust the language model according to the characteristics of sentences in the language. The experimental results show that the effectiveness of using the RNN-RBM language model for the continuous speech recognition of large vocabulary is improved to a considerable extent.
【作者单位】：湖北科技学院网络管理中心;弗吉尼亚理工大学信息技术中心;湖北科技学院外国语学院;中南财经政法大学外国语学院;
【分类号】：TN912.34

【相似文献】