基于优化肌电特征的无声语音信号识别

Unvoiced-Speech-Signal Recognition Based on Optimizing SEMG Signal Features

提出了一种基于遗传算法(GA)和fisher投影的最佳可鉴别基的求解方法.将原始特征向量向着最佳可鉴别基投影可得到具有最佳可分性的新的特征向量.从颧肌和二腹肌前腹的皮肤表面检测无声发出6个汉语元音的表面肌电信号(SEMG),以该肌电信号的AR模型系数、倒谱系数和美尔倒谱系数作为原始特征向量.使用遗传算法找出了原始特征的次优组合,并组成新的特征向量.将GA找出的次优特征向量向着fisher最佳可鉴别基投影可得到最佳鉴别特征向量.最后用改进的BP神经网络作为分类器得到了较好的识别效果.

Based on genetic algorithms （GA） and Fisher projection, a method is proposed to solve the optimal discriminant basis. New optimal eigenvector separability is obtained by projecting the original eigenvector to the optimal discriminant basis. The surface electromyogram （SEMG） signal of six unvoiced Chinese vowels are detected from the skin surface of zygomaticus major and anterior belly of the digastric, with AR model coefficients, cepstral coefficients and MFCC coefficients of SEMG signal taken as the original eigenvector. And the suboptimal eigenvector is found out from the original one by GA. Projecting the suboptimal eigenvector selected by GA to optimal discriminant basis, the optimal discriminant eigenvector is given. Experiments show that the improved BP neural network has preferable classification performance with optimal discriminant features.