Diverse Time-Frequency Attention Neural Network for Acoustic Echo Cancellation
Jinzhuo Yao, Hongqing Liu, Yi Zhou, Lu Gan, Junkang Yang · 2024
Acoustic echo cancellation (AEC) aims to eliminate echoes from near-end microphone signals and recover the near-end speech at the same time. In this work, we propose a Diverse Time Frequency Attention Neural Network (DTFAN) for AEC that operates in a full network-based manner. To that end, we first utilize a network aiming at aligning the features of the far-end reference signal and the near-end microphone signal. After that, the diverse time-frequency attentions capturing the intrinsic connections of the features in the time and frequency domains are developed. Since the alignment of the reference signal is conducted by the network, the requirement of traditionally pre-processing the far-end signal is avoided, and the whole network is end-to-end. The experimental results show that the proposed framework performs well and robustly on the synthetic test set and the blind test dataset compared to other recent approaches, especially in double-talk scenarios.