Information Extraction and Noisy Feature Pruning for Mandarin Speech Recognition

Gao G, Duan Z, Yang G, Li S, Yu X, Zhao X, Ruan J. Information Extraction and Noisy Feature Pruning for Mandarin Speech Recognition. J Audio Eng Soc. 2024;72(1/2):59-70. doi:10.17743/jaes.2022.0112

@article{Gao2024_22378,
  author = {Gao, Guozhi and Duan, Zhikui and Yang, Guangguang and Li, Shiren and Yu, Xinmei and Zhao, Xiaomeng and Ruan, Jinbiao},
  title = {{Information Extraction and Noisy Feature Pruning for Mandarin Speech Recognition}},
  journal = {Journal of the Audio Engineering Society},
  volume = {72},
  number = {1/2},
  pages = {59--70},
  year = {2024},
  month = jan,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2022.0112},
  url = {https://doi.org/10.17743/jaes.2022.0112}
}

TY  - JOUR
TI  - Information Extraction and Noisy Feature Pruning for Mandarin Speech Recognition
AU  - Gao, Guozhi
AU  - Duan, Zhikui
AU  - Yang, Guangguang
AU  - Li, Shiren
AU  - Yu, Xinmei
AU  - Zhao, Xiaomeng
AU  - Ruan, Jinbiao
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 72
IS  - 1/2
SP  - 59
EP  - 70
PY  - 2024
DA  - 2024/01/06
DO  - 10.17743/jaes.2022.0112
UR  - https://doi.org/10.17743/jaes.2022.0112
PB  - Audio Engineering Society
LA  - en
AB  - The Transformer network has two drawbacks in Automatic Speech Recognition (ASR) tasks. One is that the global features are mainly focused and other useful features, such as local features, are neglected. The other is that it is not robust to the noisy audio signal. In order to improve the model performance in ASR tasks, useful information extraction and noise removal are the main concerns. First, an information extraction module, abbreviated as IE module, is proposed to extract the local context information from the integration of previous layers which contain both low-level information and high-level information. Moreover, a noisy feature pruning (NFP) module is proposed to ease the negative effect caused by noisy audio. Finally, a network called EPT-Net is proposed on the basis of the integration of IE module, NFP module and the Transformer network. Empirical evaluations have been conducted mainly by using two widely used Chinese Mandarin datasets, which are Aishell-1 and HKUST. Experimental results can validate the effectiveness of EPT-Net, whose character error rate (CER) are 5.3%/5.6% of dev/test and 21.9% of dev in these two datasets respectively.
ER  -