mfccPSVM

LMS多麦克风语音降噪的主程序是lmsspdn.m(Multi-microphone noise reduction LMS voice is the main program lmsspdn.m)

声源定位程序，语音信号与房间脉冲响应进行卷积得到混响语音信号，再通过互相关时延估计得到定位(Sound source localization procedure, the speech signal and the room impulse response are convoluted to get the reverberation speech signal, and then through the correlation time delay estimation to obtain the localization)

MFCC特征提取实例MFCC特征提取实例MFCC特征提取实例MFCC特征提取实例(Examples MFCC feature extraction)

语音唤醒就是在设备待机状态下，用户说出特定的语音指令（唤醒词）使设备进入工作状态或完成某一操作。设备为了实现语音唤醒功能，就需要设备实时监听，也就是实时的录音并分析有没有唤醒词来唤醒设备。评价一个语音唤醒系统的标准有三个：唤醒正确率、误报率、功耗。一般情况下，唤醒正确率越高 ，误报率也越高。好的系统就需要唤醒率高，误报率低，功耗低。(Speech wakeup is when the device is in standby mode, the user speaks a particular voice command (wake up word), so that the device enters the work state or completes an operation. In order to realize the function of speech wakeup, the device needs real-time monitoring, that is, real-time recording and analyzing whether or not wake words are used to wake up the equipment. There are three criteria for evaluating a voice wakeup system: wakeup accuracy, false positive rate, and power consumption. Generally, the higher the correct rate of arousal, the higher the false positive rate. A good system requires high wake-up rate, low false positive rate and low power consumption.)

一个HMM的Matlab实现方法，可实现孤立词语音识别(Matlab implementation of a HMM methods, an isolated word speech recognition can be achieved)

有趣的语音识别程序，可以利用这个程序实现语音防盗门。是根据每个人不同的声调来实现的(interesting voice identification procedures, we can use this procedure voice security doors. Everyone is different tone to achieve the)

微软研究院的说话人识别工具包，包括GMM-UBM、I-Vector。其中demo_gmm_ubm_artificial.m和demo_ivector_plda_artificial.m为生成模拟特征参数进行训练与识别的教学示例，十分适合初学者学习说话人识别基础算法。具体使用方法请看内部文档。(Microsoft Research s speaker recognition toolkit, including GMM-UBM, I-Vector. Demo_gmm_ubm_artificial.m and demo_ivector_plda_artificial.m which generates an analog characteristic parameters for example teaching training and recognition, very suitable for beginners to learn the basic algorithm for speaker recognition. See the specific use of internal documents.)

使用TD-PSOLA（td-psola、psola）算法，做了一个语音合成，实现语音声调的变调，时长的变化等功能。参考praat源码，提取出其合成方法，效果不错。需要提供pitchtier、durationtier、pointprocess等准换信息文本（可由praat产生）供readfile函数读入(Use TD-PSOLA (TD-PSOLA, PSOLA algorithm, made a speech synthesis, speech tone tone sandhi, long-term changes in the function. Reference Praat source code, extract its synthetic method, the effect of the same.)

　提出了一种新的连续型隐马尔可夫模型(HMM ) 的概率密度函数, 并导出了一系列的参 数寻优迭代公式,(A new Continuous Hidden Markov Model (HMM) of the probability density function, and derived a series of parameters optimization iteration formula,)

通常语音信号在增强时会出现混响现象，演讲者为了消除背景混响，不得不频繁地偏转头部的方向，这样会造成脉冲响应的不断改变。我们结合盲解卷法和频谱消去法来提高逆滤波器的滤波效果。我们利用输入语音信号间的相关系数矩阵计算出稳定、精确的室内脉冲响应的逆滤波器，而这些输入信号无需测量室内的脉冲响应就能被观测到。逆滤波能够消除早期的反射，这些反射包含混响中的绝大部分能量。之后，用频谱消去法来抑制逆滤波后的信号的尾部混响。本方法在实际适应性方面的表现通过具体的实验进行了验证，结果表明盲解卷法和频谱消去法的结合相较于单独使用一种方法，能够提供一个更优越的演说环境。(Usually voice signal will be enhanced when the reverberation phenomenon, the speaker in order to eliminate the background reverberation, and had to frequently deflect the direction of the head, this will cause the impulse response of changing. We combine the blind deconvolution and spectral elimination method to improve the filtering effect of the inverse filter. We use the correlation coefficient matrix between the input speech signal to calculate the stable, accurate inverse filter of the indoor impulse response, and these input signals do not need to measure the room impulse response can be observed. Inverse filtering to eliminate early reflections, these reflections contains most of the energy in the reverberation. After the spectrum elimination method to suppress the tail reverberation of the signal after the inverse filtering. Performance in the actual adaptation of the method by specific experimental validation results show that the blind deconvolution and spectrum to eliminate )