Federated learning has a data heterogeneity (non-IID) problem: differences in speech habits, accents, and device environments among different users cause inconsistent data distributions across clients.