A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization
Apple researchers solve a critical fragility in federated ASR: pseudo-label drift that compounds across sequences and training rounds.
Why it matters
Semi-supervised federated learning for ASR has a fundamental stability problem—pseudo-label errors cascade across sequences and rounds, widening the gap to fully-supervised training. Apple's research isolates the two coupled design axes (teacher model selection and server-side anchor stabilization) that close this gap, with implications for privacy-preserving speech model training at scale.
The key facts
10 to knowSemi-supervised federated learning (SSFL) trains on unlabeled client data with server-side pseudo-labeling
ASR pseudo-label errors compound across output sequences AND across training rounds, causing divergence
Solution hinges on two design axes: teacher selection (which model generates pseudo-labels) and anchor design (server-side labeled-data updates that stabilize training)
Published by Apple Machine Learning research team
Addresses a known gap between semi-supervised federated and fully-supervised federated ASR performance
Semi-supervised federated learning (SSFL) for ASR
Pseudo-label error compounding across sequences and training rounds
Two design axes: teacher model selection and server-side anchor stabilization
Closes gap between semi-supervised and fully-supervised federated learning
Published by Apple Machine Learning Research
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Semi-supervised federated learning (SSFL) trains models on clients’ unlabeled data using a teacher to generate pseudo-labels, with a small labeled seed dataset on the server. Automatic Speech Recognition (ASR) is particularly fragile here: pseudo-label errors compound across the output sequence and…