nkcubert / dtln-aec

This Repostory contains the pretrained DTLN-aec model for real-time acoustic echo cancellation.

DTLN-aec

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 is among the top-five models of the challenge. 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.

@article{westhausen2020acoustic,
  title={Acoustic echo cancellation with the dual-signal transformation LSTM network},
  author={Westhausen, Nils L. and Meyer, Bernd T.},
  journal={arXiv preprint arXiv:2010.14337},
  year={2020}
}

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

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

---

Contents:

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.

---

Usage:

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.

---

This repository is still under construction.