RemixIT: Continual Self-Training of Speech Enhancement Models via Bootstrapped Remixing

We present RemixIT, a simple yet effective self-supervised method for training speech enhancement without the need of single isolated in-domain nor noise waveform. Our approach overcomes limitations previous methods which make them dependent on clean target signals and thus, sensitive to any domain mismatch between train test samples. RemixIT is based continuous self-training scheme in pre-trained teacher model out-of-domain data infers estimated pseudo-target mixtures. Then, by permuting remixing together, we generate new set bootstrapped mixtures corresponding pseudo-targets are used student network. Vice-versa, periodically refines its estimates using updated parameters latest models. Experimental results multiple datasets tasks not only show superiority our over prior approaches but also showcase that can be combined with separation as well applied towards semi-supervised unsupervised adaptation task. analysis, paired empirical evidence, sheds light inside functioning wherein keeps obtaining better performance while observing severely degraded pseudo-targets.