语音谱参数的增强双预测多级矢量量化的码本设计方法

表征语音谱参数的线性预测编码(LPC)参数被广泛用于各种语音编码算法。甚低位率语音编码算法要求使用尽可能少的位率编码语音谱参数。文章提出了语音谱参数的增强双预测多级矢量量化算法(EDPMSVQ)的码本设计方法。这种改进的多级矢量量化方法充分利用语音谱参数的短时相关和长时相关特性,采用了有记忆的多级矢量量化算法(MSVQ),对语音谱参数的每一维分别使用不同的预测系数;并且通过利用相邻语音帧间语音谱参数的强相关和弱相关的不同特点,采用了分别对应于强相关和弱相关的两个预测值集合,进一步减小了语音谱参数编码位率。增强双预测多级矢量量化方法能够实现20位的语音谱参数近似“透明”量化,同时能够使语音谱参数量化时的计算复杂度略有减少,所需的存储空间大为减少。

Design Method for Enhanced Dual Predictive Multi-Stage VQ of LPC Parameters

Linear predictive coding(LPC)parameters are widely used in various speech coding algorithms for represent-ing the spectral envelop information of speech.For very low bit rate speech coding algorithms ,it is important to quantize these parameters accurately using as few bits as possible.This paper presents a codebook design method of an enhanced dual predictive multi-stage vector quantization(EDPMSVQ)scheme for LPC parameters.This modified multi-stage vector quantization algorithm(MSVQ)uses multi-stage vector quantization with memory by different predictive coefficients for each dimension of LPC parameters for short-term correlation and long-term correlation,and two predictive coefficients set for strong correlation and weak correlation of LPC parameters of adjacent speech frames respectively,reduces the bit rate of LPC parameters.EDPMSVQ can achieve nearly transparent quantization of LPC parameters with20bits per speech frame ,and make the search complexity and the storage of the quantization of LPC parameters reduced.