论文部分内容阅读
语音编码算法中线谱频率(LSF)参数的量化极其重要。该文针对带有级间预测的线谱频率参数多级矢量量化算法,提出了一种多级码本之间的联合优化算法。每次迭代时,先将训练矢量对码字进行聚类,固定除当前级码本外的其他各级码本,利用加权均方误差最小原则更新当前级码本。直至达到一定的迭代步数或者两次迭代之间量化误差降低小于一定阈值,停止迭代。在一种300 b/s声码器上进行测试,结果表明该算法能够有效降低LSF参数的量化误差,从而提高合成语音的质量。
Quantization of line spectrum frequency (LSF) parameters in speech coding algorithms is extremely important. In this paper, a multi-level vector quantization algorithm for line spectrum frequency parameters with inter-stage prediction is proposed, and a joint optimization algorithm between multi-level codebooks is proposed. In each iteration, the training vectors are used to cluster the code words first, and then the other level codebooks except the current level code are fixed, and the current level codebook is updated by using the principle of weighted mean square error minimum. Until a certain number of iterations is reached or the quantization error reduction between two iterations is less than a certain threshold, stopping the iteration. The test on a 300 b / s vocoder shows that the proposed algorithm can effectively reduce the quantization error of LSF parameters and improve the quality of synthesized speech.