U-Shaped Transformer With Frequency-Band Aware Attention for Speech Enhancement

The state-of-the-art speech enhancement has limited performance in estimation accuracy. Recently, deep learning, the Transformer shows potential to exploit long-range dependency by self-attention. Therefore, it is introduced improve accuracy from a noise mixture. However, address computational cost issue with self-attention, axial attention option i.e., split 2D into two 1D attentions. Inspired attention, proposed method we calculate map along both time- and frequency-axis generate time frequency sub-attention maps. Moreover, different provides parallel multi-head attentions for frequency-axis. Furthermore, proven literature that lower frequency-band speech, generally, contains more desired information than higher frequency-band, aware high (HFA), low (LFA). U-shaped also first further extensive evaluations over four public datasets, confirm efficacy of method.