Optimal Ratio Mask-Driven Time-Frequency Neural Architecture for Acoustic Echo Cancellation
Wenzhe Lu, Liu Yong · 2025
Acoustic echo cancellation is a crucial technology for enhancing speech clarity and user experience. This paper proposes a bidirectional long short-term memory (BLSTM)-based neural architecture that combines time-domain convolutional feature extraction with frequency-domain long-term dependency modeling. It also proposes an optimal ratio masking method specifically designed for acoustic echo cancellation, overcoming the signal assumption constraints of existing masking approaches to precisely define near-end estimated speech output. Furthermore, the proposed framework integrates the feature learning capabilities of deep neural networks with the adaptive filtering characteristics of a frequency-domain state-block partitioned linear algorithm, establishing a collaborative cancellation mechanism. Experimental results demonstrate that the proposed algorithm enhances both echo cancellation performance and near-end speech quality, exhibiting superior effectiveness in acoustic echo cancellation.