语音模糊特征提取及码本训练算法

为克服低速率声码器因清浊音硬判决、粗判决而导致解码语音有帧过渡等不自然感的缺陷，在分析比较目前主流声码器编码算法中激励参数提取和量化算法的基础上，将模
糊数学中的隶属度概念引入语音子带清浊音描述。提出了五维的浊音隶属度矢量概念，用于精细描述语音丰富的激励信息；阐述了浊音隶属度矢量的提取算法；提出了矢量量化码本的模糊聚类与LBG级联训练算法（F-LBG：Fuzzy-LBG）；采用提取算法提取、建立了浊音隶属度码本的训练样本集，采用F-LBG训练了浊音隶属度码本；将提取算法和F-LBG法训练得到的浊音隶属度码本分别应用于正弦激励声码器、混合激励声码器和同态声码器的语音编、解码仿真。结果表明，用浊音隶属度矢量描述合成语音激励信号的算法，具有较高的准确性和较强的噪声鲁棒性。

Algorithm of Voice Fuzzy Feature Extraction and Codebook Training

In order to overcome the defect that the decoding speech was unnatural caused by hard decision and crude judgment,5-dimensional voiced membe r ship vector which was used to describe the information of excitation accurately, was brought forward on the basis of analyzing and comparing the algorithm which recently the mainstream vocoder used to extract and quantify the parameters of the excitation signal,and introducing the concept,membership,from fuzzy mathe matics;described the extracting algorithm of voiced membership vector and broug ht forward the VQ(Vector Quantization) codebook of fuzzy clustering and the LBG cascade training algo rithm;established the training sample set of voiced membership codebook by the means of using the extracting algorithm,and trained the codebook of voiced memb ership by the means of using F-LBG(Fuzzy LBG) algorithm.The results indicate that this al g orithm,using voiced membership vector to describe and synthesize the excitation signal,has a very high accuracy and strong noise robustness when it was used t o simulate the speech coding and decoding in the area of sinusoidal excitation v ocoder,mixed-excitation vocoder and the homomorphic vocoder.