Dual-Residual Transformer Network for Speech Recognition

Duan Z, Gao G, Chen J, Li S, Ruan J, Yang G, Yu X. Dual-Residual Transformer Network for Speech Recognition. J Audio Eng Soc. 2022;70(10):871-881. doi:10.17743/jaes.2022.0029

@article{Duan2022_22013,
  author = {Duan, Zhikui and Gao, Guozhi and Chen, Jiawei and Li, Shiren and Ruan, Jinbiao and Yang, Guangguang and Yu, Xinmei},
  title = {{Dual-Residual Transformer Network for Speech Recognition}},
  journal = {Journal of the Audio Engineering Society},
  volume = {70},
  number = {10},
  pages = {871--881},
  year = {2022},
  month = oct,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2022.0029},
  url = {https://doi.org/10.17743/jaes.2022.0029}
}

TY  - JOUR
TI  - Dual-Residual Transformer Network for Speech Recognition
AU  - Duan, Zhikui
AU  - Gao, Guozhi
AU  - Chen, Jiawei
AU  - Li, Shiren
AU  - Ruan, Jinbiao
AU  - Yang, Guangguang
AU  - Yu, Xinmei
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 70
IS  - 10
SP  - 871
EP  - 881
PY  - 2022
DA  - 2022/10/06
DO  - 10.17743/jaes.2022.0029
UR  - https://doi.org/10.17743/jaes.2022.0029
PB  - Audio Engineering Society
LA  - en
AB  - The Transformer, an attention-based encoder-decoder network, has recently become the prevailing model for automatic speech recognition because of its high recognition accuracy. However, the convergence speed of the Transformer is not that optimal. In order to address this problem, a structure called Dual-Residual Transformer Network (DRTNet), which has fast convergence speed, is proposed. In DRTNet, a direct path is added in the encoder and decoder layers to propagate features with the inspiration of the structure proposed in ResNet. Moreover, this architecture can also fuse features, which tends to improve the model performance. Specifically, the input of the current layer is the integration of the input and output of the previous layer. Empirical evaluation of the proposed DRTNet has been conducted on two public datasets, which are AISHELL-1 and HKUST, respectively. Experimental results on these two datasets show that DRTNet has faster convergence speed and better performance.
ER  -