一种可变阈值的基频曲线平滑搜索算法

汉语的声调曲线可以用基音频率轨迹来表示。传统的基频曲线平滑算法主要是中值平滑和线性平滑,这两种算法都不能有效去除基音频率中的连续错误点。本文提出可变阈值的基频曲线平滑搜索算法,根据基频点的分布选择不同阈值,得到更精确的基频曲线,且运算简单快速。分别基于人工神经网络和二次曲线拟合的孤立音节声调识别实验均表明,这种算法相比传统的中值平滑和线性平滑算法识别正确率明显提高,对提高聋儿计算机汉语教学效果具有重要意义。

A Searching Method of Fundamental Frequencies Smoothing with Variable Threshold

The pitch contour of Mandarin is presented by fundamental frequencies. The classical methods of fundamental frequencies smoothing, such as median smoothing and linear smoothing, could not correctly smooth the fundamental frequency contour when several successive frames of mistaken foundamental frequencies exist. In this paper, a new smoothing method is presented,which chooses different threshold values according to distribution of fundamental frequencies. This smoothing processing is simply and quickly performed. The experimental results of tone recognition of isolated mandarin syllables, based on neural network and secondary curve fitting respectively, show that the new method improves recognition correct rate, which is significant to improve mandarin computer aided language learning system for deaf children.