Real-time Streaming Wave-U-Net with Temporal Convolutions for Multichannel Speech Enhancement

TL;DR

This paper presents a real-time multi-channel speech enhancement system using a Wave-U-Net architecture with temporal convolutions and self-attention, achieving efficient inference suitable for conferencing applications.

Contribution

It introduces a novel streaming Wave-U-Net model with temporal convolutions, self-attention, and history cache mechanisms for real-time multi-channel speech enhancement.

Findings

Achieved real-time inference with 40ms chunks and a 0.4 real-time factor.

Maintained high speech enhancement quality comparable to offline methods.

Demonstrated effectiveness in the ConferencingSpeech2021 challenge.

Abstract

In this paper, we describe the work that we have done to participate in Task1 of the ConferencingSpeech2021 challenge. This task set a goal to develop the solution for multi-channel speech enhancement in a real-time manner. We propose a novel system for streaming speech enhancement. We employ Wave-U-Net architecture with temporal convolutions in encoder and decoder. We incorporate self-attention in the decoder to apply attention mask retrieved from skip-connection on features from down-blocks. We explore history cache mechanisms that work like hidden states in recurrent networks and implemented them in proposal solution. It helps us to run an inference with chunks length 40ms and Real-Time Factor 0.4 with the same precision.