This is the official repository of our work "Low-Complexity Acoustic Echo Cancellation with Neural Kalman Filtering" [arXiv]. (Accepted by ICASSP 2023)
👉 More results are shown on our Demo website.
The inference code and pre-trained model of NKF-AEC is released in the src folder. Try NKF-AEC by running:
python nkf.py -x ref.wav -y mic.wav -o res.wav
Note:
- NKF-AEC is a linear acoustic echo canceller.
- Time delay compensation (TDC) is necessary before running NKF if the time delay is significant (e.g., the ICASSP 2021 AEC challenge blind test set), which can be done by the GCC-PHAT algorithm, the audio fingerprinting technology, or the WebRtcAecm_AlignedFarend function in WebRTC. In such scenarios, just add the -a argument to the above command.
- The training data of the pre-trained model are derived from a small part of the AEC challenge corpus, which is introduced in the paper.
- The sampling rate of the audio is supposed to be 16 kHz.
If you find this repository helpful, please cite our work:
@article{
yang2022low,
title={Low-Complexity Acoustic Echo Cancellation with Neural Kalman Filtering},
author={Yang, Dong and Jiang, Fei and Wu, Wei and Fang, Xuefei and Cao, Muyong},
journal={arXiv preprint arXiv:2207.11388},
year={2022}
}