| 语音特点的提取技巧是语音辨认的症结技巧，研究特点提取不只有助于人们加深对语音和听觉体系的懂得，并且对进步语音辨认体系的辨认率和稳健性有着非常主要的感化。论文对今朝语音辨认体系中经常使用的特点：MFCC参数和LPC倒谱系数停止了商量，对它们的长处和缺陷停止了剖析，并把语音的正弦模子引入到了特点参数的提取进程中，提出了一种新的参数—正弦MFCC参数。在提取正弦MFCC参数的进程中，应用了语音的分解剖析法道理和听觉体系的遮蔽模子，以包管提掏出的正弦参数相符人的听觉特征。为了进步正弦MFCC参数的抗噪声机能，论文还提出了运用语音帧间的相干性停止噪声克制的办法。经由过程试验测试注解：在宁静情况中正弦MFCC参数与惯例MFCC参数的机能非常接近，辨认率仅相差0。7个百分点，在随机噪声情况中，正弦MFCC参数的机能优于惯例MFCC参数，表示出了很好的抗噪声机能。例如：在均匀信噪比为22dB的高斯白噪声情况中，男声和女声的孤立字辨认率分离比惯例MFCC参数进步了12。7和13。8个百分点。论文从语音的时光相干性动身，经由过程对正弦参数应用帧间峰值婚配算法，获得了基于正弦模子的语谱图，并提出了谱线特点的概念。作者运用谱线特点和短时能量等参数对清音段的段出息行了估量，并把提掏出的段长信息应用到了基于段长散布的隐马尔可夫模子（DDBHMM）中。经由过程对持续语音的辨认测试注解：参加了清音段长信息的判决后，使音节准确率进步了3。4%，拔出毛病率下降了20。2%。为了研究汉语语音的声学特点，法语论文网站，论文在正弦模子剖析的基本上，对汉语单韵母音节的前两个共振峰停止了剖析，在掌握运用多数几个重要参数的前提下停止了听觉的辨识试验，法语毕业论文，测试了基频和共振峰等参数的渺小变更对听觉的作用，获得了汉语单韵母的听觉混杂矩阵，并对汉语单韵母的频谱特点停止了剖析。 Abstract: Speech feature extraction technique is the key technique of speech recognition, the study of feature extraction is not only help people to deepen the understanding of speech and auditory system, and to improve the recognition rate and robustness of speech recognition system has a very important role. The current speech recognition system is often used in the characteristics: MFCC parameters and LPC coefficient of the frequency of the discussion, their strengths and weaknesses of the analysis, and the voice of the sinusoidal model introduced to the characteristics of the parameters of the extraction process, a new parameter - sine MFCC parameters. In the process of extracting sinusoidal MFCC parameters, using the decomposition analysis of speech and hearing system of the masking model, in order to ensure that the sinusoidal parameters mentioned in line with the human auditory characteristics. In order to improve the anti noise performance of sinusoidal MFCC parameters, the paper also proposes a method to restrain the noise from the coherence between speech frames. Through the process of testing and testing notes: in the quiet situation sine MFCC parameters and the practice of MFCC parameters of the function is very close, the recognition rate is only a difference of 0. 7 percentage points, in the case of random noise, the function of the sine MFCC parameters is better than the customary MFCC parameters, said a good anti noise performance. For example: in the uniform SNR 22dB Gaussian white noise, the male and female isolated word identification isolation rate than the conventional MFCC progress 12. 7 and 13. 8 percentage points. The start from the time coherence of speech, through the process of sinusoidal parameters using inter peak matching algorithm, based on sinusoidal model language spectrogram, and puts forward the spectral line features of the concept. Using the spectral characteristic and the short-time energy and other parameters of unvoiced segments segment ambition, estimated and put out of long information applied to the long spread of hidden Markov model (DDBHMM) based on. Through the process of continuous speech identification test: after taking part in the unvoiced segments length of information decision, the syllable accuracy progress 3. 4%, the rate of pull out of the problem decreased by 20. 2%. To research Chinese speech acoustic features, the sinusoidal model analysis on the basis of, the single vowel syllable in Chinese the first two formants was analyzed, and the master of several important parameters of the application of the majority of the stopped auditory identification test, test the pitch and formant parameters such as small changes of auditory effects, won the Chinese single vowel auditory hybrid matrix, and the Chinese single vowel spectrum characteristics are analyzed. 目录: 摘要 3-4 Abstract 4-5 目录 6-9 第1章 引言 9-15 1.1 语音识别的历史与近况 9-11 1.2 语音识别系统的原理和当前探讨的热点 11-12 1.3 论文的选题背景及探讨目标 12-14 1.4 论文的结构安排 14 1.5 汉语语音识别系统的测试指标说明 14-15 第2章 常用语音特征综述 15-23 2.1 语音信号的产生模型 15-17 2.1.1 经典模型 15-16 2.1.2 多带激励模型 16-17 2.2 语音信号的 LPCC（线性预测系数倒谱）参数 17-20 2.2.1 LPC 参数的基本概念及特点 17-18 2.2.2 用 LPC 法提取语音倒谱特征的基本过程 18-19 2.2.3 LPC 倒谱的特点和不足 19-20 2.3 语音信号的 MFCC 参数 20-22 2.3.1 美阶 20 2.3.2 MFCC 参数 20-22 2.3.3 MFCC 参数的特点及存在的问题 22 2.4 小结 22-23 第3章 基于正弦模型的语音特征提取 23-68 3.1 正弦模型的连续域表示 23-25 3.2 离散时间信号的正弦模型参数求解 25 3.3 本文所使用的正弦模型的参数估计法 25-29 3.4 用正弦模型参数重建语音及其客观评价指标 29-30 3.5 基于正弦模型的语音识别特征参数 30-33 3.6 听觉系统的掩蔽模型 33-38 3.6.1 听觉掩蔽特性 33-34 3.6.2 掩蔽阈值的计算 34-36 3.6.3 全局掩蔽阈值的计算 36-38 3.7 基于正弦模型的 MFCC 系数及其实验结果 38-43 3.7.1 基于正弦模型的 MFCC 特征参数 38-39 3.7.2 孤立字语音识别实验结果 39-43 3.8 正弦模型参数的抗噪性能探讨 43-51 3.8.1 随机噪声的特性略论 43-45 3.8.2 帧与帧之间的峰值匹配算法 45-48 3.8.3 利用帧间相关信息的噪声消除法 48-51 3.9 基于正弦模型参数的加噪语音识别的实验结果 51-66 3.9.1 孤立字语音的识别结果 51-61 3.9.2 连续语音的识别结果 61-66 3.10 小结 66-68 第4章 语音的正弦模型在段长信息提取方面的运用 68-78 4.1 段长信息在语音识别系统中的影响 68-70 4.2 谱线特征的定义 70-72 4.3 利用谱线特征和短时能量进行浊音段长的估计 72-74 4.4 利用浊音段长信息的连续语音识别结果 74-76 4.5 小结 76-78 第5章 汉语单韵母语音的听觉辨识 78-90 5.1 ABS 措施的优势 78-79 5.2 汉语单韵母的频谱特征及听觉辨识实验 79-85 5.3 正弦成分的幅度和相位信息对语音信号的作用 85-89 5.4 小结 89-90 第6章 结论 90-92 参考文献 92-94 致谢 94-95 个人简历、在学期间发表的学术论文与探讨成果 95 |
