FitHuBERT

This repository is for supplementing the paper, "FitHuBERT: Going Thinner and Deeper for Knowledge Distillation of Speech Self-Supervised Learning", INTERSPEECH 2022.

Distillation

1. Download the model checkpoint to perform knowledge distillation (e.g. HuBERT Base):

2. Download the LibriSpeech dataset.

Modify the configuration file in /data/conf/. The configuration file fithubert.yaml contains all the settings for reproducing FitHuBERT. Set the path to the teacher model checkpoint at teacher_model, and the root path to the LibriSpeech dataset at libri_root.

Then, run the following command:

python train.py --config ./data/conf/fithubert.yaml

After training, the model checkpoints and the corresponding configuration file will be created at /results/pretrain/.

Using the model for downstream tasks

1. Download and install the S3PRL toolkit.

2. Copy the fithubert folder into s3prl/upstream/.

3. Run the following the command to use the FitHuBERT model for automatic speech recognition(ASR).

python run_downstream.py -m train -n FitHuBERT-ASR -u fithubert -d asr -s last_hidden_state -k <path to .ckpt file> -g <path to .yaml file>

Refer to the SUPERB docs for more information on usage details and data preparation.

Checkpoint

For our checkpoints, check below links!

- FitHuBERT-100h

Checkpoint & yaml

- FitHuBERT-960h

Checkpoint & yaml

- FitW2V2-960h

Checkpoint & yaml

Citation

To cite our paper:

@article{lee2022fithubert,
  title={FitHuBERT: Going Thinner and Deeper for Knowledge Distillation of Speech Self-Supervised Learning},
  author={Lee, Yeonghyeon and Jang, Kangwook and Goo, Jahyun and Jung, Youngmoon and Kim, Hoirin},
  journal={arXiv preprint arXiv:2207.00555},
  year={2022}
}