🤖 AI Summary
This work addresses the demand for low-latency, low-complexity single-channel speech enhancement on resource-constrained embedded devices by proposing an improved architecture. Specifically, the original GRU in ULCNet is replaced with a lightweight FastGRNN, and a novel trainable complementary filter is introduced to mitigate state drift during long-duration audio inference. The proposed method achieves speech enhancement performance comparable to the original ULCNet while reducing model size by over 50% and decreasing average inference latency by 34%. These improvements significantly enhance deployment efficiency and practical applicability on edge hardware without compromising perceptual quality or intelligibility.
📝 Abstract
Single-channel speech enhancement algorithms are often used in resource-constrained embedded devices, where low latency and low complexity designs gain more importance. In recent years, researchers have proposed a wide variety of novel solutions to this problem. In particular, a recent deep learning model named ULCNet is among the state-of-the-art approaches in this domain. This paper proposes an adaptation of ULCNet, by replacing its GRU layers with FastGRNNs, to reduce both computational latency and complexity. Furthermore, this paper shows empirical evidence on the performance decay of FastGRNNs in long audio signals during inference due to internal state drifting, and proposes a novel approach based on a trainable complementary filter to mitigate it. The resulting model, Fast-ULCNet, performs on par with the state-of-the-art original ULCNet architecture on a speech enhancement task, while reducing its model size by more than half and decreasing its latency by 34% on average.