Time-frequency dual-domain attention for acoustic echo cancellation

Existing acoustic echo cancellation (AEC) technologies primarily focus on time-domain analysis, aiming to eliminate echo by modeling the long-range correlations of speech signals. However, these methods are limited in their ability to capture the dynamic variations in the frequency components of speech signals, thereby overlooking the significance of frequency-domain information. This paper proposes an energy distribution analysis method based on time-frequency (T-F) representation to address this issue. Introducing a dual-domain attention module (DDAM), which independently computes the local importance weights in both the frequency and time domains and multiplies these weights with the input features, accurately captures the most important time-frequency features of speech signals. In addition, the dual-domain feature enhancement block (DDFEB), which combines DDAM and convolutional layers, further enhances the multilevel representation of input features and integrates them into the encoder–decoder framework, effectively improving the representation of the time-frequency features. Experimental results show that the proposed method improves the perceptual evaluation of speech quality (PESQ) by 17.65% compared to the existing F-T-LSTM method and achieves a short-time objective intelligibility (STOI) score of 0.93. Furthermore, the proposed method increases the mean opinion score (MOS) by 0.33 compared to the existing DTLN-aec method, demonstrating its superiority in enhancing the user experience.