The unofficial Pytorch implementation of EffCRN: An Efficient Convolutional Recurrent Network for High-Performance Speech Enhancement (in 2023 Interspeech).
I am interested in lightweight audio noise reduction models. There are some uncertainties in the reproduction process of FCRN models, and the validity of the models has not been verified at present. You feel free to contact me by emailš« if you've got any problems.
numpy
torchinfo
pytorch >= 1.8.0
Model | Param.(paper) | Param.(here) | MACs/frame(paper) | MACs/frame(here) |
---|---|---|---|---|
FCRN | 5.2M | 5.3M | 631.5M | 632.8M |
FCRN15 | 875K | 927K | 61.5M | 64.0M |
EffCRN23 | 997K | 890K | 41M | 27.2M |
EffCRN23lite | 396K | 355K | 16M | 10.9M |
@INPROCEEDINGS{9054230,
author={Strake, Maximilian and Defraene, Bruno and Fluyt, Kristoff and Tirry, Wouter and Fingscheidt, Tim},
booktitle={ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={Fully Convolutional Recurrent Networks for Speech Enhancement},
year={2020},
volume={},
number={},
pages={6674-6678},
doi={10.1109/ICASSP40776.2020.9054230}
}
@INPROCEEDINGS{9914702,
author={Strake, Maximilian and Behlke, Adrian and Fingscheidt, Tim},
booktitle={2022 International Workshop on Acoustic Signal Enhancement (IWAENC)},
title={Self-Attention With Restricted Time Context And Resolution In Dnn Speech Enhancement},
year={2022},
volume={},
number={},
pages={1-5},
doi={10.1109/IWAENC53105.2022.9914702}
}