MELP低速率语音编码器在DSP上的改进与优化

Improvement and optimization of MELP speech coder by digital signal processor

在标准混合激励线性预测(MELP)语音编码算法中,对同时含有清音和浊音的混合帧会产生参数估计错误,并导致合成语音质量的下降。针对这个缺陷提出一种基于过渡帧判决算法的改进MELP模型(M-MELP),提高了参数估计的准确度,保证了合成语音的质量;并结合TMS320C6416数字信号处理器的结构特点,通过改写开源的MELP算法实现M-MELP算法,从程序结构、编译器和代码进行三级优化。相对于单一的优化,它能更好地提高程序运行的并行性和存储器的访问效率,从而在基于DSP的嵌入式系统中实现高质量的实时语音编码。

In the mixed excitation linear prediction（MELP） algorithm,frames include both the voiced and unvoiced sound result in parameter estimation errors and quality degradation for synthesized speech,which is a serious problem.In order to solve it,this paper gave a new approach to analyze and synthesize the transition segments which were called the transition frames model（M-MELP）.It could improve the accuracy of parameter estimation and guarantee the quality for synthesized speech.Moreover,it proposed many optimization technologies based on TMS320C6416 digital signal processor characteristics in M-MELP algorithm by rewriting and transplanting the C program of open source MELP to M-MELP.Those included program structure optimization,complier optimization,and code-level optimization.Compared with single optimization,the integrative optimization could improve program speed in parallel operation and effectual memory access more effectively.And experimental results indicate the optimized algorithm can operate well in real-time application.