论文部分内容阅读
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。
Linear predictive coding (LPC) parameters that characterize speech spectral parameters are widely used in various speech coding algorithms. Very low bit rate speech coding algorithms require that the speech spectral parameters be encoded with as few bit rates as possible. In this paper, we propose an improved codebook design method for speech spectral parameters of bi-predictive multi-level vector quantization (EDPMSVQ). This improved multi-level vector quantization method takes full advantage of short-term correlation and long-term correlation of speech spectral parameters, and uses a memory-based multi-level vector quantization algorithm (MSVQ) to separately use different spectral dimension parameters Prediction coefficient. By using the different features of strong correlation and weak correlation of speech spectral parameters between adjacent speech frames, two sets of prediction values corresponding to strong correlation and weak correlation are respectively adopted to further reduce the number of speech spectral parameter encoding bits rate. The enhanced dual-prediction multi-level vector quantization method can achieve approximately “transparent” quantification of the speech spectrum parameters of 20 bits while slightly reducing the computational complexity when quantizing the speech spectrum parameters and greatly reducing the storage space required.