[Paper]

This work proposes a source-free unsupervised domain adaptation approach for speech foundation models.

Our conda environment is provided via the file requirements.txt, please run the command below to install necessary packages:

pip install -r requirements.txt

Our code requires two kaldi-format data files: wav.scp and text.

• wav.scp contains a list of audio files, each line includes sample ID and absolute audio path:

  utt_1  /your-data-path/1.wav
  utt_2  /your-data-path/2.wav
  

• text contains a list of ground-truth transcriptions, each line includes sample ID and transcription:

  utt_1  i feel good
  utt_2  he is coming back
  

NOTE: each line in above two files should be paired.

Please refer to our training script finetune.sh and specify some settings:

• dataset: training data name;
• model_size: whisper model size;
• train_data: training data directory that contains files wav.scp and text;
• dev_data: development data directory that contains files wav.scp and text;

Then, please run command bash finetune.sh to start training. The model weights will be saved at runs/{dataset}_{model_size}.

Please refer to our inference script inference.sh and specify some settings:

• dataset: training data name;
• model_size: whisper model size;
• checkpoint: path of the trained model checkpoint (.pth file);
• test_data: test data directory that contains files wav.scp and text;

Please run command bash inference.sh for inference. WER results would be printed in the log.

We kindly hope you can cite our paper in your publication when using our research or code:

@article{hu2024self,
  title={Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models},
  author={Hu, Yuchen and Chen, Chen and Yang, Chao-Han Huck and Qin, Chengwei and Chen, Pin-Yu and Chng, Eng Siong and Zhang, Chao},
  journal={arXiv preprint arXiv:2405.14161},
  year={2024}
}