This repository contains python/tensorflow code to reproduce the experiments presented in our paper Nonlinear Residual Echo Suppression Using a Recurrent Neural Network.

The data loader uses the 'soundfile' package to read/write wavs:

pip install soundfile

Prior to training, you need to set up your data base containing echo recordings. In particular, you need a recording from the microphone signal (d), and a recording of the echo model (y) from the AEC. These recordings must not contain double-talk. Also, you need to reference a database to generate double-talk (i.e. WSJ0) in ./loaders/feature_generator.py For further details, see our paper.

To train the NRES model, use:

cd experiments
python3 nres.py train

To test the NRES model, use:

cd experiments
python3 nres.py test

This will write predictions to: ./predictions/

Microphone input with double-talk (d)

AEC residual (e) with non-linear recho and double-talk

Enhanced output (z)

Near-end signal/groundtruth (s)

Please cite our work as

@inproceedings{pfeifenberger2020residual,
  author={Lukas Pfeifenberger and Franz Pernkopf},
  title={{Nonlinear Residual Echo Suppression Using a Recurrent Neural Network}},
  year={2020},
  booktitle={Proc. Interspeech 2020},
  pages={3950--3954},
  doi={10.21437/Interspeech.2020-1473},
}