Discovering discrete subword units with binarized autoencoders and hidden-Markov-model encoders