Multi-Task Chinese Speech Recognition Method Based on the Squeezeformer Model

End-to-end training has emerged as a prominent trend in speech recognition, with Conformer models effectively integrating Transformer and CNN architectures. However, their complexity and high computational cost pose deployment challenges. To address these issues, we propose a multi-task Chinese speech recognition method based on the Squeezeformer model. We replace the FMCF structure in Conformer with an MF/CF structure, leveraging the convolutional module as a local Multi- Head Attention (MHA) module to enhance efficiency. Multilevel down-sampling and up-sampling using a time-series U-Net further reduce computational costs. By eliminating redundant LayerNorm layers and employing depthwise separable convolutions, we streamline the model, reduce parameters, and lower deployment costs. An Adaptor Layer is integrated into the MHSA module to mitigate the vanishing gradient problem, and a ScaleVar Layer is added to enhance flexibility. Additionally, the RealFormer module is introduced on the decoding side to improve context understanding. Combining Connectionist Temporal Classification (CTC) with attention-based encoding and decoding models for multi-task learning improves performance and accuracy. Experimental results show that the proposed method reduces the parameters on AISHELL-1 dataset by 16% and reduces the character error rate to 5.50%. At the same time, it also shows good performance on AISHELL-2 dataset. © (2025), (International Association of Engineers). All rights reserved.