zero-resource speech retrieval systems that require no language-specific training data or knowledge sources
Modern speech processing technologies rely on vast corpora of transcribed speech for training detailed acoustic models for each application language. While high-resource languages like English, Arabic, and Chinese have no shortage of the requisite training data, the overwhelming majority of the world’s 7000 languages are not served by the current solutions. We are developing zero-resource speech retrieval systems (keyword and document search) that require no language-specific training data or knowledge sources. These systems are built upon novel signal processing and unsupervised machine learning algorithms that exploit the rich linguistic structure of speech to facilitate speaker- and channel independent word detection and discovery. By simply speaking a search term of interest in any language, our systems can scour vast collections of speech audio for relevant information. This enables the delivery of modern information technologies for the developing world, where illiteracy remains a major barrier to adoption.