一种语音信号非周期性、周期性及基频检测的改进方法

Modified Detection of Aperiodicity,Periodicity and Pitch in Speech

APP方法可以准确检测语音信号中的非周期性、周期性和基频,是近年提出的一种先进检测新方法,对于语音基础研究和语音技术应用研究有重要作用。APP方法的最大优点是可以同时检测语音信号的基频周期、周期成分和非周期成分的能量比例,而最大缺点是计算代价巨大,运行时间为110倍实时,成为实际应用的最大障碍。该文在深入剖析APP方法的基础上,从原理架构和技术实现两个方面消除不合理的冗余处理,提出新的改进途径,发展成为改进的APP方法,即MAPP方法。MAPP方法不但加强了APP方法处理机制的合理性基础,改善基频检测的准确性和鲁棒性,而且提高计算效率约1个数量级,在CPU时钟频率为1.70GHz和内存为512MB的Pentium计算机上的运行时间加快到12.3倍实时。

The APP method is capable to provide excellent estimation of speech aperiodic / periodic measurement and pitch simultaneously which is useful in speech research and application. However, due to its heavy computational load, APP system is about 110 times real-time, being an extinct flaw for utilization. In this paper, a modified detection of aperiodicity, periodicity and pitch in speech （MAPP） method is presented, which maintains the merit of APP method and eliminates the redundancy of configuration and computation, rationalizing the methodology. Computer simulation shows that MAPP method maintains high accuracy and robustness and that the system is improved to 12.3 times real-time on Pentium processor with 1.70GHz CPU and 512MB RAM, speeding up about one order of magnitude.