Domain Adaptation Using Class Similarity for Robust Speech Recognition

Han Zhu, Jiangjiang Zhao, Yuling Ren, Li Wang, Pengyuan Zhang

This is the implementation of our paper accepted in Interspeech 2020 and the paper can be downloaded here.

Requirement

• The data preparation and GMM-HMM model training require Kaldi.
• The NN acoustic model training requires PyTorch-Kaldi.

Dataset

CommonVoice

• The CommonVoice dataset could be download from here.
• The train/dev/test data we used in this work could be found in the dataset/commonvoice/experiment_csv directory.
• The data prepare and gmm-hmm training could be done using kaldi/commonvoice/run.sh.

CHIME3

• We use the standard train/dev/test data split of CHIME3 dataset.
• The data prepare and gmm-hmm training could be done using kaldi/chime3/run.sh.

Experiment

• All experiments in the paper cound be conducted using pytorch-kaldi/run.sh.
• Configurations are stored in pytorch-kaldi/cfg.

Citation

@inproceedings{Zhu2020,
  author={Han Zhu and Jiangjiang Zhao and Yuling Ren and Li Wang and Pengyuan Zhang},
  title={{Domain Adaptation Using Class Similarity for Robust Speech Recognition}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={4367--4371},
  doi={10.21437/Interspeech.2020-3087},
  url={http://dx.doi.org/10.21437/Interspeech.2020-3087}
}

Contact

• zhuhan@hccl.ioa.ac.cn