Search

The acoustic input network based on the Conformer encoder has the problem of insufficient extraction of FBank speech information and missing channel feature information. This paper proposed an end-to-end method based on RepVGG-SE-Conformer for speech recognition to solve these problems. Firstly, the proposed model used the multi-branch structure of RepVGG to enhance the speech information extraction capability, and using the structural re-parameterization fused the multibranch into a single branch to reduce the computational complexity and speed up the model inference. Then, based on the squeeze-and-excitation network, the channel attention mechanism made up for the missing channel feature information to improve speech recognition accuracy. Finally, the experimental results on the public dataset Aishell-1 show that the proposed method's character error rate is reduced by 10.67% compared with Conformer, and the advancement of the method is verified. In addition, the proposed RepVGG-SE acoustic input network has good generalization ability in the end-to-end scene, which can effectively improve the overall performance of speech recognition models based on Transformer variants. [ABSTRACT FROM AUTHOR]

Recently, deep neural network model has become a hot research object in the field of speech recognition. However, the deep neural network relies on huge parameters and computational overhead, the excessively large model size also increases the difficulty of its deployment on edge devices. Aiming at the above problems, this paper proposed a lightweight speech recognition model based on Transformer. This method used depthwise separable convolution to obtain the feature information. Secondly, this method constructed a two half-step feed-forward layers, namely Macaron-N et, and introduced the lowrank matrix factorization to realize the model compression. Finally, it used a sparse attention mechanism to improve the training speed and decoding speed of the model. It tested on the Aishell-1 and aidatang_200zh datasets. The experimental results show that compared with Open-Transformer, the word error rate and real time factor of LM-Transformer decrease by 19. 8% and 32.1%, respectively. [ABSTRACT FROM AUTHOR]