Federated Representation Learning for Automatic Speech Recognition