【作者】 周慧芳；

【导师】 鲍鸿；

【作者基本信息】 广东工业大学 ， 检测技术与自动化装置， 2011， 硕士

【摘要】 语音识别(Speech Recognition)是让机器通过识别和理解过程把语音信号转变为相应的文本或命令的技术。语音识别技术经过几十年的探索和研究,已经取得了一系列突破性的进展,一些成熟的技术正逐渐应用于实际生活中,近年来,作为一种方便的人机交互手段,语音识别在消费类电子产品和工业控制等领域迅速发展。在嵌入式系统中实现一个能识别较大词汇量、具有较好实时性,并且具有比较高的识别率的语音识别系统,具有很高的理论研究及实用价值。论文首先介绍了语音识别技术的发展背景及国内外研究现状,并对研究的背景项目所要求的各项技术指标进行了介绍和分析。在深入研究语音识别过程中各关键模块(语音预处理,特征参数提取,识别算法)的基础上,对各模块的重点技术进行了研究。在语音预处理模块中,重点研究了预加重、分帧、加窗、短时能量和过零率相结合的端点检测法；在特征提取模块中,通过分析LPCC参数和MFCC参数的实现原理和差异,结合背景项目的要求,阐述选用MFCC参数的原因。在对语音识别的基本原理和三种经典的识别算法研究对比后,论文着重研究了隐马尔可夫(HMM)和基于此模型的语音识别算法,并且在传统的HMM算法基础上,针对HMM模型中状态数对HMM模型参数以及系统识别率的影响,对HMM的模型和识别算法进行了改进,并提出了双模板匹配法。论文在VC6.0平台上构建了一个MFCC参数与改进HMM算法相结合的语音识别系统,系统的识别率在90%以上,满足背景项目的性能要求。最后论文在对现有嵌入式操作系统深入研究的基础上,选用可移植性较好的嵌入式Windows CE操作系统,成功地在以ARM11为核心的开发板上定制和移植了Windows CE 6.0操作系统,通过跨平台的软件开发,成功在搭建好的嵌入式开发平台上实现了大词汇的语音识别系统。实验结果表明,通过对系统的改进,在嵌入式中能够实现一个具有识别率较高,实时性较好的大词汇量语音识别系统。更多还原

【Abstract】 Speech recognition is a technology which can auto convert the voice signal into te or commands by identifying and understanding. It has made a series of breakthrou after decades of exploration and research. As convenient means of human-compul interaction,speech recognition developed rapidly in the field of consumer electronics ai industrial control products recently. Speech recognition system with large vocabula recognition, good real-time and relatively high rates of correct is worth reseachin especially embedded speech recognition.Paper introduces the background and research status of speech recogniti technology firstly, and analyzes the technical indicators of project researched. The focus on the technology of the module based on the depth study of key modules recognition (voice preprocessing, feature parameters extraction, recognition algorithm Paper researches the pre-emphasis, framing, adding window, and endpoint detection (?) short-term energy and zero crossing rate for voice preprocessing; chooses MFCC feature parameters in the project after comparison between LPCC and MFCC.Then, paper introduces three classical recognition algorithms and analyzes t(?) hidden markov model (HMM) as the core algorithm. The number of the HMM mod state will influence the correct rate of recognition, as a result, paper improves the HM(?) model and recognition algorithm, and proposes a double templates matching. Paper bu(?) a speech recognition system using MFCC and improved HMM on the VC6.0 platfor(?) and the recognition rate is up to 90%, meeting the performance requirements of t(?) background project.Finally, paper chooses Windows CE operating system based on the research(?) existing embedded operating system, and succeeded in customizing and transplanting c the ARM 11 core developed board. Through cross-platform software developmen embedded development successfully built a good platform to achieve a large vocabulai speech recognition system under Windows CE 6.0 operating system. Experiment results show that the improvement on the system can achieve a better, real-time, larg vocabulary speech recognition system in the embedded system.更多还原