Protecting privacy in recorded conversations

Abstract

Professionals in the field of speech technology are often constrained by a lack of speech corpora that are important to their research and development activities. These corpora exist within the archives of various businesses and institutions; however, these entities are often prevented from sharing their data due to privacy rules and regulations. Efforts to "scrub" this data to make it shareable can result in data that has been either inadequately protected or data that has been rendered virtually unusable due to the loss resulting from suppression. This work attempts to address these issues by developing a scientific workflow that combines proven techniques in data privacy with controlled audio distortion resulting in corpora that have been adequately protected with minimal information loss.