Exponentiated Geometric-Mean-Balanced Mask for Underdetermined Source Separation

In deep-neural-network-based mask estimation for estimating the probability of speech presence for each time-frequency bin, mask estimation is biased because of the sparsity of the speech components in the bins. For unbiased estimation, we propose a method for simultaneously estimating a speech mask and its inverse (noise) mask and obtaining enhancement-mask from the speech and inverse masks’ geometric means after exponentiation to enable easier mask estimation. Experiments on underdetermined source separation confirmed the effectiveness of our mask-estimation method.