This repository contains an implementation of the deep learning-controlled acoustic echo cancellation algorithm that is described in the publication End-to-End Deep Learning-Based Adaptation Control for Linear Acoustic Echo Cancellation by T. Haubner, A. Brendel and W. Kellermann (see bibtex entry below for details).

1. Create and activate a virtual anaconda environment according to the provided YAML file.

2. Create your customized training and testing data sets which are named train_data.h5 and test_data.h5, respectively. Both datasets need to be HDF5 files with the entries

• u_td_tensor: Loudspeaker signal tensor
• y_td_tensor: Microphone signal tensor
• d_td_tensor: Ground-truth echo signal tensor
• s_td_tensor: Near-end speech signal tensor (is only required for test_data.h5)

of dimension num_sequences x signal_length, respectively, and located in the folder ./data. Note that exemplary HDF5 files are provided in the folder ./data. Yet, as the respective exemplary datasets contain only very limited amount of data, they are not suitable to train the DNN.

3. Choose your desired settings in ./main_train.py.

4. Train and test the algorithm by activating the conda environment and running python main_train.py. The processed data, including the averaged performance measures, will be saved in the subfolder ./data/proc_test_data.

If you use ideas or code from this work, please cite our paper:

IEEE Publication:

@ARTICLE{e2eDnnLinAec_ieee,
  author={Haubner, Thomas and Brendel, Andreas and Kellermann, Walter},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing}, 
  title={End-to-End Deep Learning-Based Adaptation Control for Linear Acoustic Echo Cancellation}, 
  year={2024},
  volume={32},
  pages={227-238},
  doi={10.1109/TASLP.2023.3325923}
  }

ArXiv Preprint:

@misc{e2eDnnLinAec_arxiv,
      title={End-To-End Deep Learning-based Adaptation Control for Linear Acoustic Echo Cancellation}, 
      author={Thomas Haubner and Andreas Brendel and Walter Kellermann},
      year={2023},
      eprint={2306.02450},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}