Interference Reduction in Music Source Separation Through Recurrent Inference with Deep Neural Network Models

Abstract

The issue of music source separation is discussed, focusing on its use in many audio processing tasks like remixing, target sound extraction, and audio quality improvement. This work addresses the problem through a recurrent inference framework that uses a pretrained deep neural network model, specifically within m a sk-based t i mefrequency d o main models. An iterative refinement me thod is introduced, where the output of the pretrained spectrogram separator network is fed back into the model repeatedly to extract the target source. Experiments with the Open-Unmix model show that, although the method achieves significant S i gnal-to-Interference Ratio (S IR) improve-ments comparable to top-performing discriminative models, the naive recurrent setup without mixture consistency correction can cause distortions, leading to lower Signal-to-Artifact Ratio (SAR) and Signal-toDistortion Ratio (SDR) scores. Different consistency correction techniques are then tested, demonstrating that optimizing via gradient descent in the ratio mask domain helps maintain stable performance.