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

Researchers developed a semi-supervised federated learning approach for automatic speech recognition that uses online pseudo-labels and server-side stabilization to reduce divergence and improve accuracy. Their method outperforms previous techniques in both in-domain and cross-domain settings by up to 20.8%. Key innovations include a transitioning teacher model and continuous server training on labeled data to stabilize learning.