Android平台孤立词语音识别算法研究与实现

发布时间：2018-08-29 10:05

【摘要】：语音识别是通过机器把语音信号理解、识别转成相应的命令或文本,并让机器执行文本或指令完成人机交互的技术。语音识别技术的研究对象是语音信号,其中涉及到众多的学科领域,如心理学、语言学、计算机科学、信号处理等等,它也是模式识别重要的分支。同时,在移动互联网,大数据蓬勃发展的时代背景下,涌现出更多基于智能终端的语音识别技术的应用,它们在工业、家电、通信、汽车电子、医疗、家庭服务、消费电子产品等各个领域扮演着重要的角色,所以研究智能终端语言识别技术,提高语音识别的准确率和实时性是十分有现实意义的。论文研究了语音识别关键技术,主要包括了如下四个方面:1、设计实现了基于Android平台的语音信息采集系统。实现了在Android平台上的语音信息采集、保存、音频数据的播放以及整个软件功能模块设计。为后期对语音信号做进一步分析打下基础。2、研究语音信号预处理的方法与技术。实现了语音信号的预滤波、预加重、分帧加窗、端点检测。滤除了语音信号中的电源噪声,环境噪声;获得了短时平稳的语音帧。从而计算出满足特征参数提取条件的语音信号。3、研究、改进了语音信号特征参数提取算法。着重研究了基于人耳听觉模型的梅尔倒谱系数(MFCC)特征参数提取算法;利用一半的幅度谱计算功率谱,利用频率解析度初始化梅尔三角滤波器组,对梅尔三角滤波器响应做对数的立方压缩实现算法改进,在一定程度上提高特征提取算法的性能。4、研究了语音识别模型。重点实现了动态时间规整(DTW)、隐马尔可夫(HMM)这两种识别算法。通过这两种算法对数字0到9汉语语音的特征参数进行训练和识别,利用多人采样进行实验,发现隐马尔可夫模型的识别率优于动态时间规整算法。实验结果表明:基于本文算法的孤立词语音识别系统的识别率达到60%左右,基本满足系统设计的要求。
[Abstract]:Speech recognition is a technology that can translate the speech signal into the corresponding command or text by the machine, and let the machine execute the text or command to complete the human-computer interaction. The research object of speech recognition technology is speech signal, which involves many disciplines, such as psychology, linguistics, computer science, signal processing and so on. It is also an important branch of pattern recognition. At the same time, under the background of the mobile Internet and big data's vigorous development, more and more applications of speech recognition technology based on intelligent terminals have emerged. They have been applied in industry, household appliances, communications, automotive electronics, medical treatment, and home service. Consumer electronics and other fields play an important role, so the research of intelligent terminal language recognition technology to improve the accuracy and real-time speech recognition is of great practical significance. In this paper, the key technologies of speech recognition are studied, including the following four aspects: 1. A speech information acquisition system based on Android platform is designed and implemented. The design of voice information acquisition, storage, audio data playing and the whole software function module on Android platform is realized. It lays a foundation for further analysis of speech signal, and studies the method and technology of speech signal preprocessing. Speech signal pre-filtering, pre-weighting, framing plus windows, endpoint detection. The power source noise and ambient noise are removed from the speech signal, and a short and stable speech frame is obtained. Thus, the speech signal. 3, which meets the condition of feature parameter extraction, is calculated, and the algorithm of feature parameter extraction for speech signal is improved. The algorithm of extracting characteristic parameters of Mel cepstrum coefficient (MFCC) based on human auditory model is studied emphatically, the power spectrum is calculated by half amplitude spectrum, and the Mel triangular filter bank is initialized by frequency resolution. The algorithm of logarithmic cubic compression of the response of the Mel triangle filter is improved, and the performance of the feature extraction algorithm is improved to a certain extent. The speech recognition model is studied. Two recognition algorithms, dynamic time warping (DTW), hidden Markov (HMM), are implemented. The two algorithms are used to train and recognize the feature parameters of the number 0 to 9 Chinese speech. The experimental results show that the recognition rate of the hidden Markov model is better than that of the dynamic time warping algorithm. The experimental results show that the recognition rate of the isolated words speech recognition system based on this algorithm is about 60%, which basically meets the requirements of the system design.
【学位授予单位】：湖南科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN912.34;TP316

【相似文献】