Speech Denoising in the Waveform Domain with Self-Attention

TL;DR

This paper introduces CleanUNet, a causal waveform-based speech denoising model utilizing self-attention and multi-resolution losses, achieving superior speech quality over existing methods.

Contribution

The paper presents a novel encoder-decoder model with self-attention for waveform speech denoising, improving over prior models in quality and effectiveness.

Findings

Outperforms state-of-the-art denoising models on multiple metrics

Uses self-attention to refine bottleneck representations

Optimized with multi-resolution spectrogram losses

Abstract

In this work, we present CleanUNet, a causal speech denoising model on the raw waveform. The proposed model is based on an encoder-decoder architecture combined with several self-attention blocks to refine its bottleneck representations, which is crucial to obtain good results. The model is optimized through a set of losses defined over both waveform and multi-resolution spectrograms. The proposed method outperforms the state-of-the-art models in terms of denoised speech quality from various objective and subjective evaluation metrics. We release our code and models at https://github.com/nvidia/cleanunet.