Double Branches and Stages Neural Network for Joint Acoustic Echo and Noise Suppression

In this letter, we propose a collaborative neural network framework for acoustic echo cancellation (AEC) tasks, where echo paths and spatial information are efficiently modeled through a two-stage subnetwork cascade to jointly repair the complex spectrum of the target speech. Specifically, we design...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE signal processing letters 2024, Vol.31, p.2150-2154
Hauptverfasser: Zhu, Taohua, Qian, Guobing
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In this letter, we propose a collaborative neural network framework for acoustic echo cancellation (AEC) tasks, where echo paths and spatial information are efficiently modeled through a two-stage subnetwork cascade to jointly repair the complex spectrum of the target speech. Specifically, we design a two-path structure consisting of a real part and an imaginary part as well as an amplitude phase, a lightweight network module to suppress part of the echo and potential noise in the first stage, and a larger network to estimate the complex residuals for phase correction and spectral restoration in the second stage. In order to maximize the performance of the model at each stage, we propose two convolutional modules: an inplace gate convolutional module and a complex squeezed temporal convolutional module (CSTCM). In addition, a cross-domain loss function is designed to improve the generalization capability. Experiments are conducted under various mismatch scenarios, and the results show that the proposed dual-path staging method provides superior performance over other advanced methods.
ISSN:1070-9908
1558-2361
DOI:10.1109/LSP.2024.3441492