Architecture of the Obs-TasNet to estimate observation filter coefficients. Additional layers compared to the original IC Conv-TasNet [37] are displayed red. Expressions in parentheses at signal lines indicate tensor size, parentheses in layers correspond to a transform of dimensions. 1D Conv refers to a standard 1D convolutional layer, TCN Conv to a TCN convolution block with varying colours for different dilation factors d, 1 x 1 Conv to a pointwise 2D convolution, and PReLU to a parametric rectified linear unit activation function.