This is the original Pytorch implementation for our paper. Our work focuses on building a noise adaptive speech enhancement technique with an incremental strategy to reduce catastrophic forgetting problem, complementing existing noise adaptation strategies without using additional storage.

• Installation
• Steps and Usages
• Citation
• Contact

In order to install SERIL, clone the repo and install it using pip or python :

git clone https://github.com/ChangLee0903/SERIL
cd SERIL
# Install install-required deps
pip install numpy Cython
# Install all the necessary packages
pip install -r requirements.txt

1. Data Preprocess:

   You can mix your own data by using adnoise function in util.py and set the data path in config/config.yaml.

   • Prepare noisy wavs directory, e.g. "../train/noisy_T0/", "../dev/noisy_T0/", "../test/noisy_T0/"
   • Prepare clean wavs directory, e.g. "../train/clean_T0/", "../dev/clean_T0/", "../test/clean_T0/"

   train: 
    noisy:
      ['../train/noisy_T0', '../train/noisy_T1', '../train/noisy_T2', '../train/noisy_T3', '../train/noisy_T4']
    clean:
      ['../train/clean_T0', '../train/clean_T1', '../train/clean_T2', '../train/clean_T3', '../train/clean_T4']
    ...
    

   Make sure that both noisy and clean directory have the same wave file names.

   ../train/noisy_T0/p226_006.wav
   ../train/clean_T0/p226_006.wav
   ...

2. Train the networks for each domain sequentially:

   Note that the utterances of first domain will be taken as the training data of pretrain task which will train much longer than other domains, and both --mode finetune and --mode seril will share the same pretrain model.

   • Check log directory
   • Choose seril/naive fine-tuning mode
   • Run:

   python main.py --logdir log --do train --mode finetune
   python main.py --logdir log --do train --mode seril
   

3. Inference:

   Make sure your --logdir log setting is the same with the training step.

   • Check log directory
   • Choose seril/naive fine-tuning mode
   • Run:

   python main.py --logdir log --do test --mode finetune
   python main.py --logdir log --do test --mode seril

If you find the code helpful in your research, please do consider cite us!

@inproceedings{SERIL,
  author={Chi-Chang Lee, Yu-Chen Lin, Hsuan-Tien Lin, Hsin-Min Wang, Yu Tsao},
  title={{SERIL:} Noise Adaptive Speech Enhancement using Regularization-based Incremental Learning},
  year=2020,
  booktitle={Proc. Interspeech},
}

Any bug report or improvement suggestion will be appreciated!

e-mail: [email protected]