This Repostory contains the pretrained DTLN-aec model for real-time acoustic echo cancellation in TF-lite format. This model was handed in to the acoustic echo cancellation challenge (AEC-Challenge) organized by Microsoft. The DTLN-aec model reached the 3rd place. The results of the AEC-Challenge can be found here.

The model was trained on data from the DNS-Challenge and the AEC-Challenge reposetories.

The arXiv preprint can be found here. Please cite:

@INPROCEEDINGS{westhausen21_dtln_aec,
  author={Westhausen, Nils L. and Meyer, Bernd T.},
  booktitle={ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={{Acoustic Echo Cancellation with the Dual-Signal Transformation LSTM Network}}, 
  year={2021},
  volume={},
  number={},
  pages={7138-7142},
  doi={10.1109/ICASSP39728.2021.9413510}
  }

Author: Nils L. Westhausen (Communication Acoustics , Carl von Ossietzky University, Oldenburg, Germany)

This code is licensed under the terms of the MIT license.

This repository contains three prtrained models of different size:

• dtln_aec_128 (model with 128 LSTM units per layer, 1.8M parameters)
• dtln_aec_256 (model with 256 LSTM units per layer, 3.9M parameters)
• dtln_aec_512 (model with 512 LSTM units per layer, 10.4M parameters)

The dtln_aec_512 was handed in to the challenge.

First install the depencies from requirements.txt

Afterwards the model can be tested with:

$ python run_aec.py -i /folder/with/input/files -o /target/folder/ -m ./pretrained_models/dtln_aec_512

Files for testing can be found in the AEC-Challenge respository. The convention for file names is *_mic.wav for the near-end microphone signals and *_lpb.wav for the far-end microphone or loopback signals. The folder audio_samples contains one audio sample for each condition. The *_processed.wav files are created by the dtln_aec_512 model.