仿人听觉系统的发育网络模型研究

发布时间：2018-04-14 04:11

本文选题：听觉模型 + 发育网络　；参考：《郑州大学》2017年硕士论文

【摘要】：随着人工智能技术的普及,机器人已逐渐进入人类的日常生活。让机器人听懂人类的语言,就是使机器人能够像人类一样获得听觉。目前,声音识别是机器听觉领域的核心方向,相比于机器听觉,人类听觉系统能准确、快速地识别声音的内容,所以模拟人类听觉系统构建人工听觉模型已成为机器听觉领域研究的热点。本文引入发育网络,通过粗略模拟人类听觉系统信息处理的腹侧通路结构来构建听觉模型,腹侧通路主要处理声音的内容信息,即主要负责声音的内容识别。为了体现所构建模型的性能,本文设置语音识别实验进行验证。本文的工作主要分为三部分。首先简单概述了人类听觉系统的结构,并对其信息传导通路进行详细的分析,为构建听觉模型提供理论基础。接着重点介绍了脑壳封闭的发育网络的基础理论,对自主发育的思想进行了阐述,并解释了发育网络的基本模型和学习算法。其次,基于发育网络,粗略模拟人类听觉系统信息处理的腹侧通路结构建立发育网络模型。简述了发育网络模型的特点,并对其模型结构提出改进。对发育网络引入多种仿生机制,如设计神经元的连接方式和学习状态实现神经元的自主学习;采用神经元再生机制实现神经元的充分利用;增加神经元释放机制解决神经元的资源分配问题。设计了发育网络模型的发育程序,并录制语音创建语音库,然后设置了语音识别实验对发育网络模型的性能进行验证。实验结果表明,发育网络模型对英文单词和短语的识别率分别可达到94.44%和90.19%,说明模型初步具备了对实际环境中语音的识别能力。最后,为了进一步提高发育网络模型的性能,引入了深度信念网络构建特征提取器对语音提取深度特征,分析了语音经过深度信念网络的特征表示,并对其特征进行可视化。在原有语音数据库的基础上,结合新特征,对发育网络模型进行语音识别实验。从实验结果可以看出,发育网络模型在新特征的基础上对英文单词和短语识别率分别可达96.30%和92.96%,表明模型的性能得到了较大的改善。与传统的语音识别模型相比,发育网络模型表现出较优的性能。
[Abstract]:With the popularization of artificial intelligence technology, robots have gradually entered the daily life of human beings.To make robots understand human language is to enable robots to get hearing just like humans.At present, sound recognition is the core direction in the field of machine hearing. Compared with machine hearing, human auditory system can recognize the content of sound accurately and quickly.Therefore, simulating human auditory system to build artificial auditory model has become a hot spot in the field of machine hearing.In this paper, the developmental network is introduced to construct the auditory model by roughly simulating the ventral pathway structure of human auditory system information processing. The ventral pathway mainly deals with the content information of sound, that is, it is mainly responsible for the content recognition of sound.In order to reflect the performance of the model, this paper set up speech recognition experiments to verify.The work of this paper is divided into three parts.Firstly, the structure of human auditory system is briefly summarized, and its information transmission pathway is analyzed in detail, which provides a theoretical basis for the construction of auditory model.Then the basic theory of the closed development network of the skull is introduced, the idea of autonomous development is expounded, and the basic model and learning algorithm of the development network are explained.Secondly, the development network model is established based on the development network, which roughly simulates the ventral pathway structure of human auditory system information processing.In this paper, the characteristics of the developmental network model are briefly described, and the improvement of its model structure is put forward.A variety of biomimetic mechanisms are introduced to the development network, such as designing the connection mode and learning state of neurons to realize autonomous learning of neurons, adopting the mechanism of neuron regeneration to make full use of neurons.The mechanism of neuronal release is added to solve the problem of resource allocation of neurons.The development program of the development network model is designed and the speech database is recorded. Then the performance of the development network model is verified by setting up the speech recognition experiment.The experimental results show that the developmental network model has a recognition rate of 94.44% and 90.19% for English words and phrases respectively.Finally, in order to further improve the performance of the development network model, a feature extractor constructed by the depth belief network is introduced to extract the depth features of speech, and the feature representation of speech passing through the depth belief network is analyzed, and the features are visualized.Based on the original speech database and the new features, the speech recognition experiment of the development network model is carried out.It can be seen from the experimental results that the recognition rates of English words and phrases are 96.30% and 92.96%, respectively, based on the new features of the developmental network model, which indicates that the performance of the model has been greatly improved.Compared with the traditional speech recognition model, the developmental network model has better performance.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP242

【相似文献】