Speech Retrieval in Unknown Languages: a Pilot Study

Most cross-lingual speech retrieval assumes intensive knowledge about all involved languages. However, such resource may not exist for some less popular languages. Some applications call for speech retrieval in unknown languages. In this work, we leverage on a quasi-language-independent subword recognizer trained on multiple languages, to obtain an abstracted representation of speech data in an unknown language. Languageindependent query expansion is achieved either by allowing a wide lattice output for an audio query, or by taking advantage of distinctive features in speech articulation to propose subwords most similar to the given subwords in a query. We propose using a retrieval model based on finite state machines for fuzzy matching of speech sound patterns, and further for speech retrieval. A pilot study of speech retrieval in unknown languages is presented, using English, Spanish and Russian as training languages, and Croatian as the unknown target language.