A Correspondence Variational Autoencoder for Unsupervised Acoustic Word Embeddings