Fig. 3
From: Explicit-memory multiresolution adaptive framework for speech and music separation
Visual of the proposed memory informed attentional gating for a single sample. The attention module takes a single batch of representations (H) as the input and reshapes it to (q, (p.r)). Utilizing reshaped \(\tilde{H}\) and the selected object memory vector \(O_x\), we estimate the anchor memory \(R_x\). In the next step, we gate the embeddings that do not fire coherently with the Anchor Memory and produce the gated representation \(\hat{H}\)