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

This post introduces a practical recipe for Semi-Supervised Federated ASR (SSFL) by focusing on two coupled design axes: the teacher (model generating pseudo-labels) and the anchor (server-side updates stabilizing training). It demonstrates that per-client online teachers, once stabilized, match or beat broadcast global teachers. It also highlights the critical role of server-side training on labeled data between rounds for convergence, and the sensitivity of this stabilization to data augmentation and batch size. The findings provide guidelines for SSFL in ASR, improving over prior methods.