End-to-end Speech Recognition Method Based on Conformer

To enhance the extraction of local features while preserving the contextual extraction ability of the Conformer encoder, an end-to-end speech recognition method called SEResConformer is proposed. This method can effectively extract speech feature information during the acoustic input stage. First, the residual structure of ResNet is utilized to perform deep convolution on the speech feature matrix. Then, missing information is supplemented through shortcut connections. During this process, the audio features are weighted by the SEnet module. Finally, the feature information processed by ResNet and the feature information processed by the channel attention mechanism are fused to improve the accuracy of speech recognition. Experimental results on the public dataset Aishell-1 demonstrate that the proposed method achieves a 7.1% reduction in character error rate compared to the original Conformer, verifying the feasibility of the approach and reducing the computational cost.