Fast-ULCNet: A Fast and Ultra Low-Complexity Network for Single Channel Speech Enhancement

Abstract

Single-channel speech enhancement algorithms are often used in resource-constrained embedded devices, where low latency and low-complexity designs are essential. We propose Fast-ULCNet, an adaptation of ULCNet that replaces GRU layers with FastGRNNs, reducing computational latency and complexity. To address state drifting in long signals, we introduce a trainable complementary filter. Fast-ULCNet matches state-of-the-art performance while cutting model size by over 50% and reducing latency by 34% on average.

Audio Examples

This demonstration presents a collection of audio examples designed to allow visitors to evaluate and compare different noise reduction techniques. All the clips have been chosen from the public test set available on the 2020 DNS Challenge repository. For each audio sample, we provide the original noisy recording alongside two enhanced versions: one processed with the state-of-the-art model ULCNet, and the other processed with our proposed method, Fast-ULCNet.

Conditions of Use

1. All material used for training and testing is publicly available on the DNS Challenge GitHub repository.
2. Samples are for testing and demonstration of noise reduction techniques only.
3. No warranties are expressed or implied.
4. Do not redistribute or modify outside of evaluation without permission.
5. Include this notice when sharing links to this page.