一种改进的MELP算法

An improved MELP algorithm

介绍了混合激励线性预测(MELP)的算法,并研究了MELP的特点.在MELP的基础上,针对其在语音信号不稳定的部分包含不规则的基因周期,它们的自相关函数值比较小,因而引入噪声的问题,采用了滑动窗的方法来减少或避免上述问题.此外还针对LSF量化的过程中,其码本的存储量与计算的复杂度大的问题采用了3级矢量量化的方法,从而可以把MELP的码率降到2.1 kbps左右,仍有较好的合成语音质量.

Mixed excited linear prediction （MELP） algorithm is introduced, and its characteristics is studied. In previous MELP algorithm, since the voice signal in response to the instability of the gene contains the irregular cycles, their autocorrelation function is relatively small, thus it can cause noise problems. Therefore a sliding window method is proposed to reduce or eliminate these problems. In addition, in consideration of the storage capacity of the code table and complex computation in the process of LSF quantification, a three-level vector quantization meth- od is utilized, which drops the code rate to around 2.1 kbps, but still has good synthetic voice quality.