摘 要

随着科学技术的发展，语音识别技术(Automatic Speech Recognition，ASR)已经渐渐融入到人们生产生活的各个领域中，如工业、医药、电子设备、通信、家电等应用方向。由于语音信号具有很大的时变性、连续性，所以对连续语音进行识别具有很大的难度，而关键词识别是从连续语音中检测出若干关键词，只需对用户所关心的信息进行识别，并不需要将所有的字符识别出来，大大降低了语音识别的难度。因此，关键词语音识别更具有实际应用意义。本文主要采用DTW算法和HMM模型进行关键词语音识别技术仿真分析研究。

本文的主要研究内容如下：

（1）明确关键词语音识别的目的及意义，阅读参考文献了解语音识别技术的国内外发展及研究现状。

（2）讨论分析语音信号的产生过程及其产生的数学模型。

（3）分析语音的短时加窗、短时能量、短时平均过零率等语音特征，分析讨论端点检测和Mel特征参数提取的相关原理，并以女生数字1汉语发音为例对语音信号的各声学特征进行了详细的论述分析。

（4）在基于特征分析的基础上，详细分析并讨论了语音识别的广泛应用的两种方法的相关原理，即DTW算法和HMM模型，分别采用这两种方法对数字0~9的汉语发音进行仿真分析。

（5）基于MATLAB R2014a平台设计一个简单的固定（7字长）电话号码识别系统，从时间上对采样得到的连续语音进行分割，分别采用DTW算法和HMM模型对分割后得到的关键词语音进行识别仿真，将得到的识别结果进行比较分析，并进一步阐述了这两种算法的特点和优劣势。

关键词：语音识别；DTW；HMM；端点检测；特征提取

Abstract

With the development of technology, Automatic Speech Recognition (ASR) technology has gradually permeated various fields of people production and living, such as industry, medicine, electronics, communication, household appliances and other application direction. Because of the voice signal has character of non-stationary and continuity, it is so difficult to deal with continuous speech recognition, but keywords recognition just to identify several keywords detected from continuous speech, only get the key information that users care about, which need not to identify the whole sentence, greatly reduce the difficulty. Therefore, the speech recognition of keywords is more practical. This paper mainly uses the DTW algorithm and the HMM model to perform the keyword speech recognition technology simulation and analysis.

The main work of this article is as follows:

(1) To be sure purpose and meaning of the speech recognition and to realize the development and research status of speech recognition technology by reading relative document literature.

(2) This paper discusses the production mechanism of speech signal and the mathematical model.

(3) Analyzed short-term window, short-term energy, short-term average zero crossing rate, such as these phonetic characteristics, analyzed the endpoint detection, and the related principle of Mel characteristic parameters extraction, and put the girl's number 1 Chinese pronunciation as an example to the acoustic characteristics of speech signals are analyzed in detail.

(4) Based on the characteristic analysis, clearly analyzed and discussed the wide application of speech recognition of the related principle of two methods, the DTW algorithm and the HMM model, these two methods were used respectively to the numbers 0 to 9 Chinese pronunciation for simulation analysis.

(5) To design a simple fixed (7 characters in length) telephone number identification system by using MATLAB R2014a platform,divided continuous speech by time to get continuous speech segmentation, respectively, by using the DTW (dynamic time warping) algorithm and the HMM model for keywords that speech recognition simulation after segmentation, will receive the identification results of comparative analysis, and further expounds the characteristics and disadvantages of these two kinds of algorithm.

Keyword: speech recognition; DTW; HMM; endpoint detection; characteristic extraction