LAPE

We introduce LAPE, a unified framework for Low-resource Audio Processing and Evaluation.

What is LAPE?

LAPE is an easy-to-use toolkit for audio processing. In its initial release, LAPE supports Self-Supervised Learning (SSL)-based Upstream Pre-training and Downstream Fine-tuning. LAPE, originally introduced in this paper, integrates all our research on low-resource audio processing in one unified framework. We open-source LAPE to promote more research in this space.

How to use?

Setup

conda create -n lape -y python=3.8
conda activate lape
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install -r requirements.txt

Upstream SSL Pre-training

Its simple! First create a csv file with a single column named files. That column should have absolute paths to your raw wav audio files. Next, for upstream SSL pre-training using un-labeled data just run:

python train_upstream.py --input /path/to/csv/ --upstream name_of_upstream

The name_of_upstream should be the name of the upstream algorithm you want to use. The upstreams we currently support are: Additionally, all other defaults (encoder, learning rate, etc.) are set in the upstream specific configs. An example can be seen in slicer/config.yaml. Feel free to change it according to your needs!

Downstream Task Fine-tuning

Once you have pre-trained your encoder using SSL, fine-tune it for any task using this command:

python train_downstream.py --input name_of_task --ckpt /path/to/pretrained/ckpt

If the task is supported by HuggingFace🤗 datasets library, the script automatically downloads the data. If not, you need to specify additional arguments including --train_csv, --valid_csv and --test_csv. The csvs should have 2 columns including wav_path and label where the former is the path to the raw wav file and the latter is the tag of the audio. Similar to upstream, we maintain a config file for all other defaults. An example can be seen in src/downstream/downstream_config.yaml.

Also feel free to remove the --ckpt argument if you want to train an encoder on your downstream encoder from scratch (no SSL pre-training).

Contribution Guidelines

If you want to contribute (models or algorithms or a completely new feature!) please feel free to open an issue followed by a pull-request.

Found a bug?

Please raise an issue and we will try our best to resolve it as-soon-as-possible!

Cite

If you find this toolkit useful, please consider citing following papers.

If you have used the LAPE toolkit in your experiments, or the DeLoRes-S or DeLoRes-M SSL pre-training algorithms:

@ARTICLE{9868132,
  author={Ghosh, Sreyan and Seth, Ashish and Umesh, S},
  journal={IEEE Journal of Selected Topics in Signal Processing}, 
  title={Decorrelating Feature Spaces for Learning General-Purpose Audio Representations}, 
  year={2022},
  volume={16},
  number={6},
  pages={1402-1414},
  doi={10.1109/JSTSP.2022.3202093}}

If you have used the MAST audio encoder for your work:

@inproceedings{ghosh2023mast,
  title={MAST: Multiscale Audio Spectrogram Transformers},
  author={Ghosh, Sreyan and Seth, Ashish and Umesh, S and Manocha, Dinesh},
  booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2023},
  organization={IEEE}
}

For SLICER, UNFUSED and DECAR:

@inproceedings{seth2023slicer,
  title={SLICER: Learning universal audio representations using low-resource self-supervised pre-training},
  author={Seth, Ashish and Ghosh, Sreyan and Umesh, S and Manocha, Dinesh},
  booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2023},
  organization={IEEE}
}

@inproceedings{seth2023unfused,
  title={UNFUSED: UNsupervised Finetuning Using SElf supervised Distillation},
  author={Seth, Ashish and Ghosh, Sreyan and Umesh, S and Manocha, Dinesh},
  booktitle={ICASSP 2023-2023 SASB Workshop},
  year={2023},
  organization={IEEE}
}

@article{ghosh2021deep,
  title={Deep clustering for general-purpose audio representations},
  author={Ghosh, Sreyan and Katta, Sandesh V and Seth, Ashish and Umesh, Srinivasan},
  journal={arXiv preprint arXiv:2110.08895},
  year={2021}
}

Contact and Contributors

Ashish Seth (email: cs20s030@smail.iitm.ac.in) Sreyan Ghosh (email: sreyang@umd.edu)