基于深度神经网络的蒙古语语音识别系统声学模型的研究

发布时间：2018-01-09 10:14

本文关键词：基于深度神经网络的蒙古语语音识别系统声学模型的研究　出处：《内蒙古大学》2017年硕士论文　论文类型：学位论文

【摘要】：语音识别是实现人机交互的重要手段,它的最终目的是让人与机器能够像人与人一样通过语音对话实现交流。在最近几年里,语音识别技术经历了巨大的变革,深度神经网络成功替代了高斯混合模型(GMM)成为声学建模的主流方法。不同的神经网络结构如深度神经网络(DNN)、卷积神经网络(CNN)以及长短时记忆神经网络(LSTM)等在声学模型上得到了广泛的研究。然而这些研究主要集中在英语,中文等使用人数较多的语言上,对于其它使用人数较少的语言来讲,语音识别技术研究仍然处于初级的研究阶段,其中蒙古语正是这样的一种语言。本文将DNN、时延神经网络(TDNN)、CNN,LSTM,前向序列记忆神经网络(FSMN)等神经网络结构用于蒙古语声学模型的建模,通过对比不同神经网络结构的声学模型来研究神经网络结构对蒙古语语音识别性能影响。同时为了进一步提升蒙古语声学模型的性能,本文还使用神经网络的区分性训练方法和添加说话人特征两种方式对模型进行了优化训练。实验结果表明,基于LSTM的声学模型在蒙古语语音识别系统中有最好的性能,但是LSTM由于结构复杂导致解码过程计算量大于其他结构,而FSMN在准确率和计算复杂度的均衡性能上要优于其他结构。同时使用区分性训练和添加说话人特征后,声学模型的性能有明显的提升。
[Abstract]:Speech recognition is an important means to achieve human-computer interaction, its ultimate aim is to let people and machines like people like to communicate through speech dialogue. In recent years, speech recognition technology has undergone great changes, the depth of the neural network has successfully replaced the Gauss mixture model (GMM) has become the mainstream method of acoustic modeling neural network. Various structures such as the depth of the neural network (DNN), convolutional neural network (CNN) and the short memory neural network (LSTM) in the acoustic model has been widely studied. However, these studies mainly focus on English, such as the use of more Chinese number of language, for other use fewer words in terms of speech recognition technology is still in the primary stage of the study, which is a kind of Mongolian language such as DNN. In this paper, time delay neural network (TDNN), CNN, LSTM, prior to the sequence of memory of God The network (FSMN), the structure of neural network modeling for Mongolian acoustic model, by comparing the acoustic model of different kinds of neural network to study the influence of neural network structure on the performance of Mongolian speech recognition. At the same time in order to further enhance the performance of Mongolian acoustic models, discriminative training methods are used in this neural network and add speaker characteristics in two ways optimize the training of the model. The experimental results show that the acoustic model based on LSTM has the best performance in the Mongolian speech recognition system, but due to the LSTM complex structure lead to calculate the amount of decoding process is greater than the other structures, and the computation complexity FSMN equalization performance is superior to other structures in the accuracy and use of distinction. At the same time add training and speaker characteristics, performance of acoustic model has significantly improved.

【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TN912.34

【相似文献】