This is an Unofficial PyTorch implementation of the paper "NSNet2: Data augmentation and loss normalization for deep noise suppression" (https://arxiv.org/abs/2008.06412).
The model is based on the NSNet2 baseline model from the DNS-Challenge. The pipeline is based on the paper "Towards efficient models for real-time deep noise suppression" (https://arxiv.org/abs/2101.09249).
The model/nsnet2 model structure is same as the original model.
The model/nsnet2_ex model is a modified version of the original model. Includes the preprocessing and postprocessing steps in the model, but excludes the FFT and IFFT processes.
In addition, the model/tinyNSNet model is a modified version of the original model. The model is implemented with a small number of parameters via tensorflow. Replace the original model GRU with a FastGRNN cell. The quantized tf-lite model get 0.067 ms inference time on Apple M2 chip.
| Model | Platform | Inference Time |
|---|---|---|
| tinyNSNet | Apple M2 | 0.067 ms |
Pretrained model NSNet2 by Microsoft is licensed under CC BY 4.0
The baseline NSNet noise suppression:
@misc{braun2020data,
title={Data augmentation and loss normalization for deep noise suppression},
author={Sebastian Braun and Ivan Tashev},
year={2020},
eprint={2008.06412},
archivePrefix={arXiv},
primaryClass={eess.AS}
}