Implicit Neural Spatial Filtering for Multichannel Source Separation in the Waveform Domain

TL;DR

This paper introduces an end-to-end neural model for multichannel source separation that implicitly performs spatial filtering directly in the waveform domain, effectively separating moving sound sources without traditional spatial features.

Contribution

The proposed model is a novel single-stage causal waveform-to-waveform neural network that performs implicit spatial filtering for dynamic multichannel source separation.

Findings

Matches performance of oracle beamformer plus single-channel enhancement

Effective in dynamic acoustic scenes with moving sources

Operates without traditional spatial features or processing components

Abstract

We present a single-stage casual waveform-to-waveform multichannel model that can separate moving sound sources based on their broad spatial locations in a dynamic acoustic scene. We divide the scene into two spatial regions containing, respectively, the target and the interfering sound sources. The model is trained end-to-end and performs spatial processing implicitly, without any components based on traditional processing or use of hand-crafted spatial features. We evaluate the proposed model on a real-world dataset and show that the model matches the performance of an oracle beamformer followed by a state-of-the-art single-channel enhancement network.