causal-speech-enhancement

Project repo for Deep Learning 02456 at DTU

4 nov

• [x] todos on slack
• [x] hpc init
• [ ] and training
• [ ] downsizing data
• [ ] 1-2 weeks -> setup data and training loop, reduce loss, gap bw two models
• [ ] then tinker

11 Nov

• [ ] Resampling, downsampling -> why and how?
• [ ] Loss functions
• [ ] Conv TAS Net
• try batch size = 1
• torch compile if too slow
• wav2vec is open source - on outputs- good starting poit -similar to what they did before
• so we need to basically figure out a good way to do the transfer learning

References

EARS_DATASET

@inproceedings{richter2024ears,
  title={{EARS}: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation},
  author={Richter, Julius and Wu, Yi-Chiao and Krenn, Steven and Welker, Simon and Lay, Bunlong and Watanabe, Shinjii and Richard, Alexander and Gerkmann, Timo},
  booktitle={ISCA Interspeech},
  year={2024}
}

Reference Links:

1. HPC Related: https://www.hpc.dtu.dk/?page_id=59#scratch https://www.hpc.dtu.dk/?page_id=927#Scratch_usage https://www.hpc.dtu.dk/?page_id=2534 https://www.gbar.dtu.dk/index.php/faq https://www.hpc.dtu.dk/?page_id=2501 https://www.hpc.dtu.dk/?page_id=4317 https://www.hpc.dtu.dk/?page_id=59#scratch https://www.hpc.dtu.dk/?page_id=927#Scratch_usage https://arxiv.org/abs/1809.07454

2. Dataset: https://github.com/sp-uhh/ears_benchmark https://pytorch.org/docs/stable/data.html#torch.utils.data.Dataset

3. Model Architecture Related: https://github.com/pytorch/audio/blob/main/src/torchaudio/models/conv_tasnet.py https://docs.python.org/3/library/glob.html

4. Literature: https://arxiv.org/pdf/2408.11842v2 https://arxiv.org/abs/1809.07454

Aastha says: Kill me already >.<