Towards Personalized Federated Learning for Dysarthric Speech Recognition

This paper introduces personalized federated learning (FL) strategies for automatic speech recognition (ASR) tailored to dysarthric speakers, addressing the challenges posed by speaker variability. It evaluates two aggregation methods—parameter-based averaging and embedding-based averaging—demonstrating statistically significant improvements in word error rate (WER) over the baseline regularized FedAvg, achieving reductions of up to 0.99% on UASpeech and 0.56% on TORGO. These findings highlight the potential of personalized FL in enhancing ASR performance for diverse speaker profiles, which is crucial for developing more effective and inclusive speech recognition systems.