Acoustic Echo Cancellation Network Based on Harmonic-aware Attention

Existing adaptive filtering and deep learning-based AEC techniques struggle with echo suppression and capturing speech harmonic structures in dynamic environments. to address this, the study proposes HL-ELD, an echo cancellation network featuring the harmonic-aware attention module (HLAM) that fuses local and global attention for better spectral feature response. the Harmonic-aware Global Block (HGB) enhances time-frequency modeling using multi-step convolutional and LSTM networks. the network employs a convolutional encoder for feature extraction, a bidirectional LSTM for temporal enhancement, and a decoder for signal reconstruction. experiments show significant improvements in speech naturalness and coherence in clean and noisy environments.