Multilingual acoustic word embedding models for processing zero-resource languages

Acoustic word embeddings are fixed-dimensional representations of\nvariable-length speech segments. In settings where unlabelled speech is the\nonly available resource, such embeddings can be used in "zero-resource" speech\nsearch, indexing and discovery systems. Here we propose to train a single\nsupervised embedding model on labelled data from multiple well-resourced\nlanguages and then apply it to unseen zero-resource languages. For this\ntransfer learning approach, we consider two multilingual recurrent neural\nnetwork models: a discriminative classifier trained on the joint vocabularies\nof all training languages, and a correspondence autoencoder trained to\nreconstruct word pairs. We test these using a word discrimination task on six\ntarget zero-resource languages. When trained on seven well-resourced languages,\nboth models perform similarly and outperform unsupervised models trained on the\nzero-resource languages. With just a single training language, the second model\nworks better, but performance depends more on the particular training--testing\nlanguage pair.\n