Researchers at Apple Machine Learning have introduced a new approach to semi-supervised federated learning (SSFL) specifically designed for Automatic Speech Recognition (ASR). The method aims to close the performance gap between semi-supervised federated learning and fully-supervised federated learning, which is often caused by error compounding in pseudo-labels during training rounds.
NewsAppleApple Machine Learning
Apple Researchers Propose Method to Stabilize Semi-Supervised Federated ASR via Online Pseudo-Labels
The proposed technique focuses on two key design axes: the "teacher" and the "anchor." The "teacher" refers to the model that generates pseudo-labels—unlabeled data marked with predicted correct outputs. The study shows that using a per-client online teacher, which evolves alongside the client's own model, can match or outperform a broadcast global teacher, particularly in-domain and under domain shifts. The "anchor" refers to the server-side updates on labeled seed data. The researchers found that continuous training on labeled data at the server is essential to prevent the online teacher from drifting.
The integration of these two axes significantly improves performance. According to the research, the method outperformed the strongest prior methods on 9 out of 11 test pairs, achieving an average improvement of 20.8% in-domain and 10.0% in cross-domain scenarios.
Sources
- A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization （Apple Machine Learning, 2026-09-24）
- arXiv