MKConvLFormer: A streaming speech recognition algorithm for power grid semantic environments

Abstract The power grid dispatching system is a critical component in the development of the power industry and the technological innovation of smart grids. However, as smart grid systems expand, the traditional approach to managing and dispatching the grid through telephone communication is no longer sufficient to meet the growing data processing demands of smart grids. Consequently, this paper introduces an algorithm named MKConvLFormer (multi-scale residual convolutional neural network and Linformer). The MKConvLFormer is a streaming speech recognition algorithm designed for the semantic environment of the power grid, which utilizes the transformer model to extract features and incorporates a multi-scale convolutional neural network within the transformer architecture to enhance its feature extraction capabilities. Additionally, this research leverages the low-rank property of matrices to reduce the computational complexity of the existing self-attention mechanism, aiming for efficient and accurate recognition of streaming speech within the power grid semantic environment. Experimental results indicate that the MKConvLFormer model demonstrates superior speech recognition performance on the AISHELL-1 and THCHS-30 datasets.