基于VMD的双通道构音障碍语音特征图谱提取算法

VMD Based Binary Channels Speech Feature Map Extraction Algorithm for Dysarthria

针对在提取构音障碍患者语音有效特征信息不足,导致语音识别率低的问题,提出一种基于变分模态分解(VMD)的多尺度双通道滤波器组(MBCFbank)特征图谱提取算法.首先,为了更好地提取符合人耳听觉结构特性的声学特征,提出一种双通道滤波器组(BCFbank)特征提取算法,该算法采用Mel滤波后做对数变换,同时采用Gammatone滤波后作非线性响度变换;其次,采用VMD来优化BCFbank特征,对分解后的多个语音信号分量筛选出相关系数较高的3个,分别提取其BCFbank特征及其差分特征,同时对未分解的语音信号提取BCFbank特征,从而构成MBCFbank特征图谱;最后,在双路语音识别模型上进行训练和识别.实验结果表明,基于BCFbank特征、MBCFbank特征图谱的语音识别模型准确率最高分别达到了87.82%,94.34%,优于Fbank特征的识别效果.

A multiscale binary channels filter banks(MBCFbank)feature extraction algorithm based on variational modal decomposition(VMD)is proposed to address the issue of poor speech recognition caused by insufficient extraction of effective feature information from speech of patients with dysarthria.Firstly,in order to better extract the acoustic features that conform to the structural characteristics of human ears,a binary‑channels filter banks(BCFbank)feature extraction algorithm is proposed,which uses Mel filtering and performs logarithmic transformation,simultaneously using Gammatone filtering to perform nonlinear loudness transformation.Secondly,VMD is used to optimize the BCFbank features.Three components with higher correlation coefficients are selected from the decomposed multiple speech signal components,and their BCFbank features and differential features are extracted respectively.At the same time,BCFbank features are extracted from the undecomposed speech signals to form the MBCFbank feature map spectrum.Finally,training and recognition are conducted on a dual channel speech recognition model.The experimental results show that the speech recognition model based on BCFbank features and MBCFbank feature maps has the highest accuracy of 87.82%and 94.34%,respectively,which is superior to the recognition effect of Fbank features.