Research on Additive Margin Softmax Speaker Recognition Based on Convolutional and Gated Recurrent Neural Networks

Lan C, Wang Y, Zhang L, Zhao H. Research on Additive Margin Softmax Speaker Recognition Based on Convolutional and Gated Recurrent Neural Networks. J Audio Eng Soc. 2022;70(7/8):611-620. doi:10.17743/jaes.2022.0018

@article{Lan2022_21827,
  author = {Lan, Chaofeng and Wang, Yuqiao and Zhang, Lei and Zhao, Hongyun},
  title = {{Research on Additive Margin Softmax Speaker Recognition Based on Convolutional and Gated Recurrent Neural Networks}},
  journal = {Journal of the Audio Engineering Society},
  volume = {70},
  number = {7/8},
  pages = {611--620},
  year = {2022},
  month = jul,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2022.0018},
  url = {https://doi.org/10.17743/jaes.2022.0018}
}

TY  - JOUR
TI  - Research on Additive Margin Softmax Speaker Recognition Based on Convolutional and Gated Recurrent Neural Networks
AU  - Lan, Chaofeng
AU  - Wang, Yuqiao
AU  - Zhang, Lei
AU  - Zhao, Hongyun
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 70
IS  - 7/8
SP  - 611
EP  - 620
PY  - 2022
DA  - 2022/07/06
DO  - 10.17743/jaes.2022.0018
UR  - https://doi.org/10.17743/jaes.2022.0018
PB  - Audio Engineering Society
LA  - en
AB  - To improve the recognition rate of the speaker recognition system, a model scheme combined with the Additive Margin--Softmax loss function is proposed from the perspective of model differentiation and based on the fusion of Convolutional Neural Network and Gated Recurrent Unit, which not only reduces the distance of similar sample features and increases the distance among different types of sample features simultaneously but also uses layer normalization to constrain the distribution of high-dimensional features. In order to address the problem of poor robustness of the speaker recognition system in real scenes, the SpecAugment data enhancement method is proposed to train the speaker model to combat external environmental interference. Based on the experimental data, the speech recognition performance of the proposed and traditional methods is analyzed. The experimental results show that, compared with other models, the equal error rate based on the Additive Margin--Convolutional Neural Network--Gated Recurrent Unit method is 4.48%, and the recognition rate is 99.18%. Adding layer normalization to the training model can improve the training speed to a certain extent, and the speaker model has better robustness.
ER  -