基于神经网络的声音识别算法研究

发布时间：2018-04-29 16:34

本文选题：声音识别 + 神经网络　；参考：《北京邮电大学》2014年硕士论文

【摘要】：随着大数据时代的到来,无论在工业生产还是日常生活环境中都充斥着大量的多媒体数据,而声音作为多媒体数据中的重要组成部分蕴含了大量的信息。对声音数据进行处理和分析可以从大量的数据中挖掘出对我们有用的信息,因此针对声音的处理和分析技术一直以来都是各国学者深入研究的热点。其中声音识别技术近些年来也得到了大量的关注和应用。声音识别是将待识别声音的特征与声音样本特征进行比对,从而得到待测声音和样本的一致性判断。声音识别可应用于许多领域和场合,如环境声音异常监测、音频资料检索、音频媒体版权监测等。在对声音进行识别之前先需要对其进行前期处理,声音识别前期的处理流程包括预加重、分帧加窗和端点检测等。在前期处理的基础上,对声音进行特征提取得到声音的特征向量。接着是模式匹配阶段,通过模式匹配得到声音识别的最终结果。基于神经网络的基本工作原理,本文主要研究了如何应用神经网络解决多类声音识别中的模式匹配问题。本论文的主要工作如下： 1,在介绍了神经网络基础知识的基础上,针对两类识别网络的具体参数进行了研究探讨,确定了传输函数、神经元和神经层个数等多个参数。 2,探讨了多类声音识别的识别方案,对线型识别、并行排名、两类晋级三种不同的识别方案进行对比论证,确定了以两类晋级识别为基础的多类声音识别方案。 3,详细阐述了运用两类识别神经网络对多类声音进行识别的方法。对多组竞争方法和可信率进行了全面的阐述,通过多组竞争的方法可以大幅提高两类神经网络的识别率,具体实例验证了多组竞争方法在多类识别中的效用。将可信率的计算应用到识别程序中可以让用户主动掌握识别进程,得到在识别时间和识别率之间权衡后满意的识别结果。 4,针对声音类别总数是任意数的情况,论述了分组匹配竞争方法。通过多个具体实例讨论了分组匹配识别方法的基本规律,总结了类别数十以内的声音识别推荐分组模型,类别数更大的问题可以通过先分组到十以内的小组来解决。
[Abstract]:With the arrival of big data era, both industrial production and daily life environment are full of a lot of multimedia data, and sound as an important part of multimedia data contains a lot of information. The processing and analysis of sound data can extract useful information from a large number of data, so the technology of sound processing and analysis has always been the focus of deep research by scholars all over the world. In recent years, sound recognition technology has also been a lot of attention and application. Sound recognition is to compare the features of the sound to be identified with the characteristics of the sound samples, so as to obtain the consistency of the sound and the samples to be tested. Sound recognition can be used in many fields and applications, such as environmental sound anomaly monitoring, audio data retrieval, audio media copyright monitoring and so on. The pre-processing of voice recognition is needed before it is recognized. The pre-processing process includes pre-weighting, framing and endpoint detection. On the basis of previous processing, the feature vector of sound is obtained by feature extraction. Then there is the pattern matching stage, and the final result of sound recognition is obtained by pattern matching. Based on the basic working principle of neural network, this paper mainly studies how to use neural network to solve the problem of pattern matching in multi-class sound recognition. The main work of this thesis is as follows: 1. On the basis of introducing the basic knowledge of neural network, the specific parameters of two kinds of recognition networks are studied and discussed, and several parameters, such as transmission function, number of neurons and neural layers, are determined. 2. The recognition schemes of multi-class sound recognition are discussed, and the linear recognition, parallel ranking and two kinds of promotion three different recognition schemes are compared and proved, and the multi-class sound recognition scheme based on the two kinds of promotion recognition is determined. 3. Two kinds of recognition neural networks are used to recognize multi-class sound in detail. In this paper, the method of multi-group competition and the probability of trustworthiness are expounded. The recognition rate of two kinds of neural networks can be greatly improved by the method of multi-group competition. The effectiveness of multi-group competition method in multi-class recognition is verified by an example. The application of the trust rate calculation to the recognition program can enable the user to master the identification process actively and obtain satisfactory recognition results after balancing the recognition time and the recognition rate. 4. In view of the situation that the total number of sound categories is arbitrary, the competition method of grouping matching is discussed. The basic rules of grouping matching recognition are discussed through several concrete examples, and the recommended grouping model of sound recognition within tens of categories is summarized. The problem of larger number of categories can be solved by grouping into groups within ten first.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP18;TN912.34

【参考文献】