硕士学位论文
孤立词语音识别
算法的研究和实现
THE RESEARCH AND
IMPLEMENTATIONOF ALGORITHM OFISOLATED WORD SPEECH RECOGNITION
李硕新员工培养方案
哈尔滨工业大学
2010年12月国内图书分类号:TM431.2 学校代码:10213
国际图书分类号:621.3 密级：公开硕士学位论文
孤立词语音识别
算法的研究和实现
硕士研究生: 李硕
导师: 王明江教授
申请学位: 工学硕士
学科专业: 微电子学与固体电子学
所在单位: 深圳研究生院电视连续剧红娘子
答辩日期: 2010 年 12 月
授予学位单位: 哈尔滨工业大学 Classified Index: TM431.2
U.D.C: 621.3Dissertation for the Master Degree of Engineering
THE RESEARCH AND
IMPLEMENTATIONOF ALGORITHM OFISOLATED WORD SPEECH
RECOGNITIONCandidate: Lishuo
Supervisor: Prof. Wang Mingjiang
Academic Degree Applied for: Master of Engineering
Microelectronics and Solid-State
Specialty:
Electronics
Affiliation:Shenzhen Graduate School
Date of Defence: December, 2010
Degree-Conferring-Institution: Harbin Institute of Technology 哈尔滨工业大学工学硕士学位论文
摘要
首都休闲大学
语音识别技术以语音信号处理为研究对象。本文主要研究小词汇量、非特定人、孤立词的汉语语音识别算法与实现。
文章首先介绍了隐马尔可夫模型HMM，包括 HMM 的参数估计,Viterbi
算法等。接着阐述了如何利用 HMM 构建语音识别系统，分别讨论了基于离散HMMDHMM和连续 HMMCHMM的孤立词语音识别系统。 CHMM 和 DHMM
的差别在于观测值的概率分布函数。DHMM 的概率分布函数是离散的概率值, 而对 CHMM 则是连续的概率密度,CHMM 昀常采用的概率密度函数是高斯混
合模型GMM。DHMM 具有运算量小、存储少的优点，但由于 DHMM 存在量
化误差，识别精度较 CHMM 低。
在孤立词语音识别系统中，词表外OOV的语音输入将会对语音识别系统产
生难以预料的结果，这是设计中需要避免的，因此研究拒识算法变得十分重要。
拒识算法主要有两种方法，一是利用废料模型，二是利用已有的识别结果进行
辨识。第一种方法需要额外训练废料语音，而且一般难以选择合适的废料语音。
因此实际应用中一般选择第二种方法。本文通过研究识别结果，昀终通过识别
概率中的昀大和次大概率差作为拒识算法的判别依据。
在 PC机上难以模拟在实际设备上所使用的语音识别系统，例如背景噪声的干扰，语音信号的实时输入和实时处理，本文介绍了在 TI 公司的 DM642 DSP撞月
历史首位!拜登得票数突破8000万
上实现孤立词语音识别的方法，包括语音信号的输入，以及 VAD算法的实现等。
关键词: 孤立词；语音识别；隐马尔可夫；拒识算法
-I- 哈尔滨工业大学工学硕士学位论文
Abstract
The research object in Speech recognition technology is based on speech signal
processing. This paper studies the algorithm and implementation of small vocabulary,
speaker-independent, Chinese words, isolated word recognitionThe article first introduced the Hidden Markov Model HMM, including HMM parameter estimation, Viterbi algorithm. Then explains how to build speech
recognition system using HMM.HMM were discussed based on discrete HMM DHMM and continuous HMM CHMM.The main difference between CHMM and DHMM is that the probability distribution function of observations. DHMM’ s
probability distribution function of observations is discrete probability, while the
CHMM is a continuous probability density, In commonly CHMM uses Gaussian
mixture model GMM as the probability density function. DHMM has the advantages of the small computational load, the less of storage, but the existence of
quantization error will lead to the lower recognition accuracy compare to CHMMIn the practical application of speech recognition systems, the speech of out of屈服强度
vocabulary OOV input on voice recognition system will cause unpredictable results,
which is need to avoid in the design,so studying rejection algorithm become very
important. In general, the realization of rejection algorithm has two ways, one is to
use garbage model, the second is the use of the existing results of the recognition. The
first method requires additional training data, and generally difficult to select the
appropriate data of garbage. Therefore, in general application we choose the second
method. This paper studied the results of recognition, and ultimately chosen the
difference between largest and second largest between the recognition probability as
the basis of rejection algorithmIt is difficult to simulate the practical application environment of the speech
recognition systems, such as the interference of background noise, voice, real-time
input and real-time signal processing. This paper introduced the implementation of
isolated speech recognition on TI' s DM642 DSP ,including how to input the speech
豫公网安备41160202000603
站长QQ:729038198
关于我们
投诉建议