浊音隶属度参数及F-LBG算法

为了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中,提出了5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;介绍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);用提取算法提取、建立了浊音隶属度码本的训练样本集,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进行语音编、解码仿真;结果表明,用浊音隶属度矢量描述和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。

Voiced membership parameters and F-LBG algorithm.

In order to overcome the defect of unnaturalness caused by hard decisions of voiced/unvoiced speech segments in LPC based low bit rate vocoder, a 5-dimensional voiced membership vector for accurately describing the information of excitation is proposed on the basis of analyzing and comparing the algorithms which are used by today＇ s mainstream vocoders for extracting and quantifying parameters of the excitation signal. This paper introduces the concept of member- ship from fuzzy mathematics, describes the extracting algorithm of voiced membership Vector and the VQ codebook of fuzzy clustering, and the LBG cascade training. Computer simulation experiments are conducted. The results indicate that this algorithm, using voiced membership vector to describe and synthesize the excitation signal, can achieve a very high accuracy and strong noise robustness when it is used to simulate the speech coding and decoding process for sinusoidal ex- citation vocoder, mixed-excitation vocoder and the homomorphic vocoder.