A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization

Apple researchers have introduced a method to improve semi-supervised federated learning for automatic speech recognition by using server-side update stabilization to address error compounding in pseudo-labels. This approach allows models to be trained on large amounts of unlabeled client data while maintaining accuracy, which is historically difficult due to the fragility of speech recognition systems during decentralized training. The technique stabilizes the learning process by integrating a small, labeled server-side dataset to guide the teacher models that generate labels for client devices.