个数可变脉冲线性预测编码研究

不同语音帧的激励信号复杂性不同,所以采用相同个数的脉冲作为激励信号并不合理。针对这一点,提出了个数可变脉冲线性预测编码算法。该算法不固定脉冲个数,而是根据激励信号的复杂度而确定。个数可变脉冲线性预测编码的目的是用尽量少的脉冲数来满足误差约束,这可以看作一个稀疏表示问题。进而,给出了具体的脉冲搜索算法以及个数可变脉冲线性预测编码方案。实验结果发现增加脉冲可以减少误差,但是前面搜索出的脉冲对误差的贡献要大于后搜索出的脉冲。与G.723.1和G.729比较发现,个数可变脉冲线性预测编码可以在约4.2 kbps的编码速率下获得优于G.723.1的合成语音,但略差于G.729。本文算法的编码时间较长,是下一步需要解决的问题。

Number-variable pulse linear prediction coding

It is not reasonable to use the same number of pulses for all frames with different complexity. For this, number-variable pulse linear prediction coding (NVP-LPC) is presented in this paper. In this algorithm, the number of pulses is not fixed but determined according to the complexity of speech signal. NVP-LPC aims at satisfying the error constraint at the cost of the least pulses, which can be seen as a sparse representation problem. Moreover, the detailed pulse search algorithm and coding scheme are presented. Results show that adding pulses can reduce the error, and the pulses determined earlier make more contributions to reduce the error than latter pulses. Comparing with G.723.1 and G.729, results show NVP-LPC can obtain the synthesized speech superior to G.723.1 but slightly worse than G. 729 at the coding rate of about 4.4 kbps.