Residual Echo Suppression Using Dual-Stream Interactive Transformers With Selective Multi-Scale Encoding

Traditional acoustic echo cancellation (AEC) employs linear adaptive filters to identify the echo path between the speaker and the microphone. Although AEC significantly enhances audio quality in voice communications, residual echo persists due to factors such as inaccurate estimation of echo path induced by nonlinear distortion of the far-end signal. Consequently, a post-suppression module is essential for achieving sufficient echo attenuation. This paper proposes a novel time domain end-to-end method with selective multi-scale encoder and attentional interactive module for nonlinear residual echo suppression (RES) in double-talk scenarios. Specifically, the selective multi-scale encoder can adaptively assign the channel-wise weights to features with multiple time resolutions based on the changing acoustic environment, thereby regulating the feature stream to effectively recover near-end speech. Moreover, the attentional interactive module provides a novel context-aware fusion strategy. Differing from conventional linear fusion operations, such as addition and concatenation employed in existing RES methods, this module dynamically calculates fusion weights for dual signal streams, enabling the neural network to benefit from their correlations. Experimental results demonstrate the superiority of the proposed method.