A multiscale feature extraction algorithm for dysarthric speech recognition

Abstract

In this paper, we propose a multi-scale mel domain feature map extraction algorithm to solve the problem that the speech recognition rate of dysarthria is difficult to improve. We used the empirical mode decomposition method to decompose speech signals and extracted Fbank features and their first-order differences for each of the three effective components to construct a new feature map, which could capture details in the frequency domain. Secondly, due to the problems of effective feature loss and high computational complexity in the training process of single channel neural network, we proposed a speech recognition network model in this paper. Finally, training and decoding were performed on the public UA-Speech dataset. The experimental results showed that the accuracy of the speech recognition model of this method reached 92.77%. Therefore, the algorithm proposed in this paper can effectively improve the speech recognition rate of dysarthria.