A novel speaker normalization method based on formant recovery and mellin transform 基于子波变换的自适应滤波语音增强方法
4.
Formant : it uses the estimation of the overall shape of the spectrum ' s amplitude envelope 共振峰:使用了对频谱的振幅包络线的大体形状所作的估计。
5.
The results of simulation indicated speech signal processed by the optimum algorithm presents obvious periodicity in time domain , and effect of the formant is removed or restrained effectively in frequency domain 处理后的语音信号在时域上表现出明显的周期性特征，同时在频域上也观察到声道的共振峰结构影响得到消除或有效的抑制。
6.
Carrying the information of content , it has the function of content verification . this article compares various method of formant parameters calculation , and completes the calculation of the 5 former formant frequencies and bands . as pitch frequency 在分别使用支持向量回归和多项式回归进行句子转换的对比实验中,支持向量回归的频谱更接近目标语音,但其执行速度慢,运算时间较长。
7.
Moreover in speech enhancement , especially in reducing the pulse noise , morphological algorithm has its unique advantage . particularly morphological filter may maintain the preferable accurate of the speech signal in speech waveform , and which produces little impairment to the formant of speech . so the spectrum structure of the speech is retained well , and the quality of the speech will not be reduced 特别是，在时域波形分析中，形态学滤波增强较小波去噪更好地保持语音信号的细节；在频域分析中，形态学滤波对语音信号的基音频率、频谱斜率、共振峰等语音特征的影响很小，因而能够较好的保留语音信号的频谱结构，使语音品质不致降低。
8.
This thesis tries to update the cmdsr system to achieve the characters below : real - time , better robust , higher recognition rate , non - special - man . considering the disadvantages of traditional improved spectrum subtraction speech enhancement , this thesis proposes the theory of fuzzy spectrum subtraction based on the fuzzy theory and improved spectrum subtraction speech enhancement ; as for the difficulties of detecting the endpoint of speech signal , the thesis gives the table of initial and the improved parameters , with which we can confirm the endpoints of mandarin digit speech ; the thesis puts forward two - level digit real - time speech recognition system , the first level is based on discrete hidden markov model which is linear predictive coding cepstrum ( lpcc ) and difference linear predictive coding cepstrum ( dlpcc ) , the second level is based on formant parameters ; as for the realization of hardware , the thesis depicts the realization of every part of cmdsr based on the tms320vc5402 in detail ; as for the development of software , the thesis gives the software design flow chart of cmdsr , simulates the basic theory with matlab language and gives the simulation results 针对传统的“改进谱相减法语音增强”参数设定单一、环境适应能力差的缺点,提出了一种利用模糊理论和“改进的谱相减法”结合的“模糊谱相减法语音增强” ;针对语音信号端点检测困难的特点,通过matlab仿真试验,给出了能够准确确定数码语音端点的初始和改进参数表;提出了利用基于线性预测编码倒谱参数和差分线性预测编码倒谱参数相结合的离散隐含马尔可夫模型进行第一级识别、利用共振峰参数进行第二级识别的两级汉语数码语音识别系统,在保证系统实时性的同时,实现连接汉语数码语音识别系统识别率的提高;在硬件实现上,详细阐述了基于tms320vc5402的连接汉语数码语音识别系统各部分硬件设计;在软件开发上,给出了连接汉语数码语音识别的软件设计各部分的流程图,并对各部分进行了matlab仿真,并给出了仿真结果。
9.
In various speech character parameters , formant frequency , bandwidth and pitch frequency are chosen as voice character parameters . the reasons are as follows : hearing apperceive experiments indicates that formant frequency can stand for a majority of voice information , while average pitch frequency can explain 55 % ability of speaker verification 数据结果与多项式回归和线性多变量回归相比,支持向量回归既提高了泛化性能又避免了频谱不连续性,从而使转换语音与目标语音的频谱距离失真分别减少了33 . 29 %和35 . 24 % 。
10.
Firstly , we study the construction of emotion - speech template database , and analyze the common features such as pitch , energy and formant . after choosing the useful features by using fuzzy entropy effectiveness analysis , we get better performance with the application of neural network . in addition , we propose some more efficient features such as speech rate , pitch slope , mel - frequency cepstral coefficients and its transient parameters , and design a processing model based on vector quantization for cepstral features to fusing different features 本文首先介绍了情感语音数据库的建立情况,然后研究了基音频率、振幅能量和共振峰等目前常用的情感特征在语音情感识别中的作用,并且通过一种基于模糊熵的特征有效性分析方法进行了有效特征的筛选,应用人工神经网络建立了初步的语音情感识别模型,经过实验发现特征筛选后系统的识别效果有着一定程度的提高。
