Fast-ULCNet

Official repository of Fast-ULCNet.

The paper is available here.

A demo with online examples is available here.

This repository contains the code to build the Comfi-FastGRNN and Fast-ULCNet model in Tensorflow 2+ and Pytorch.

The Comfi-FastGRNN layer is available as a pip package, making it easy to integrate into any TensorFlow or PyTorch model.

---

Installation

Clone the repository, create an environment and install the dependencies.

Install requirements

pip install -r requirements.txt

Install Comfi-FastGRNN Tensorflow layer

pip install comfi-fast-grnn-tensorflow 

Install Comfi-FastGRNN Pytorch layer

pip install comfi-fast-grnn-torch

Build model

Tensorflow

python fast_ulcnet_networks/tensorflow_version/FastULCNet.py

Pytorch

python fast_ulcnet_networks/pytorch_version/FastULCNetTorch.py

Unit test

A simple unit test code is provided to compare the Comfi-FastGRNN implementations between Tensorflow and Pytorch.

python fast_ulcnet_networks/unit_tests/unit_test_tensorflow_torch.py

To do list

• Fast-ULCNet Pytorch implementation
• Python package of Comfi-FastGRNN for both Tensorflow and Pytorch

Citation

If you use Fast-ULCNet to inspire your research, please cite the paper:

@INPROCEEDINGS{11463365,
  author={Larraza, Nicolás Arrieta and de Koeijer, Niels},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Fast-ULCNet: A Fast and Ultra Low Complexity Network for Single-Channel Speech Enhancement}, 
  year={2026},
  volume={},
  number={},
  pages={16822-16826},
  keywords={Filtering;Filters;Circuits and systems;Media Access Control;Protocols;HTTP;Speech codecs;Instant messaging;Modulation;Network architecture;deep learning;speech enhancement;low complexity;low latency},
  doi={10.1109/ICASSP55912.2026.11463365}}