基于机器学习的数字音频分类方法研究.docVIP

基于机器学习的数字音频分类方法研究 摘要： 让计算机能听懂人类的语言，是人类自计算机诞生以来梦寐以求的想法。随着经济的发展，人们越来越迫切要求摆脱键盘的束缚而代之以语音输入这样便于使用的、自然的、人性化的输入方式。这就使语音信号分析成为社会生活中越来越重要的一部分。其中语音识别是语音信号分析的重要领域，而语音特征信号识别又是语音识别研究领域的一个重要方面。 语音特征信号识别一般采用模式匹配的方法解。首先通过预处理提取语音特征，作为该语音片段的模型。将该模型与已知的参考模型相比较，获得最佳匹配的参考模式作为识别结果，在论文中研究的是基于传统的BP神经网络的音频分类与SVM技术音频分类的仿真实验结果进行比较，对比两者各自优缺点，从而熟悉这两种技术的基本工作原理和算法。通过实验对其性能进行了主观评价和客观数据分析，对于所选语音信号BP网络具有较快的训练速度，但是SVM技术具有较高的识别率。BP网络的学习和记忆具有不稳定性。也就是说，如果增加了学习样本，训练好的网络就需要从头开始训练，对于以前的权值和阈值是没有记忆的。但是可以将预测、分类或聚类做的比较好的权值保存。关键词：支持向量机esearch on digital audio classification base on machine-learning Abstract: Since the birth of the computer people want to let the computer can understand humans language .People are becoming more and more urgent to get rid of keyboard and replace it by voice input that is personalized、convenient and natural to be used .So that Analysis of speech signal become a more and more important part in social life.The speech recognition is an important part of Analysis of speech signal ,and the voice signal recognition is one of the most important aspects of speech recognition . Speech feature signals recognition use to use pattern matching method.Firstly, through pretreatment of phonetic feature extraction make as the model.The model is compared with known reference model, get the best matches the reference pattern as a result of recognition. In this paper research is based on the traditional BP neural network audio classification and technology of Support vector machine audio classification the experimental results were compared to find each of advantages and disadvantages, so to understand with the two technical basic principle and algorithm. Through the experiments on the performance of the subjective evaluation and objective data analysis, for selected speech signal BP network has faster training speed, but the SVM technology has higher recognition rate. BP network learning and memory have instability. That is to say that if the increase of learning samples, the trained