极低速率语音编码中LSP参数的高效量化算法  

Efficient LSP quantization algorithm for very low bit rate speech coding

作　　者：李军林[1] 崔慧娟[1] 唐昆[1] 

机构地区：[1]清华大学电子工程系微波与数字通信技术国家重点实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2004年第10期1422-1425,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(69972020)

摘　　要：为在极低速率下实现高质量的语音编码,提出了一种新的有效的线谱对(LSP)参数量化算法——P-RS-MSMQ算法。此算法以多帧联合矩阵量化作为基本框架,引入了基于超级帧模式的均值去除和帧间预测策略、矩阵分裂和子矩阵多级量化策略;同时提出了基于语音帧短时谱能量的帧内加权和基于超级帧中各子帧重要性的帧间加权策略等。实验表明:此算法能够在700b/s的速率下获得接近透明量化的性能;即使在300～400b/s的极低速率下也具有较高质量的量化效果。因此该算法的实现对极低速率语音编码算法的研究具有重要的意义。This paper presents an efficient linear spectrum pairs (LSP) parameters quantization algorithm called Pattern-based Residual Split Multi-Stage Matrix Quantization (P-RS-MSMQ). The method uses matrix quantization of multipleframes with not only mean-value-removal and inter-frame prediction basedon thesuper-frame pattern, but also matrix splitting and multi-stage matrix quantization. The algorithm also uses an intra-frame weighting scheme based on the energy of the short-time spectrum and an inter-frame weighting scheme based on the importance offrames in the super-frame. Simulation results demonstrate that the algorithm can achieve almost 'transparent' LSP quantization at 700 b/s and good performance at the very low bit-rateof 300～400 b/s , so the algorithm is very useful for very low bit-rate speech coding.

关 键 词：语音编码 线谱对 超级帧 矩阵量化 

分 类 号：TN912.3[电子电信—通信与信息系统]