Federated Self-training for Semi-supervised Audio Recognition

Federated Learning is a distributed machine learning paradigm dealing with decentralized and personal datasets. Since data reside on devices such as smartphones virtual assistants, labeling entrusted to the clients or labels are extracted in an automated way. Specifically, case of audio data, acquiring semantic annotations can be prohibitively expensive time-consuming. As result, abundance remains unlabeled unexploited users’ devices. Most existing federated approaches focus supervised without harnessing data. In this work, we study problem semi-supervised models via self-training conjunction learning. We propose FedSTAR exploit large-scale on-device improve generalization recognition models. further demonstrate that self-supervised pre-trained accelerate training models, significantly improving convergence within fewer rounds. conduct experiments diverse public classification datasets investigate performance our under varying percentages labeled Notably, show little 3% available, average rate by 13.28% compared fully model.