APPROACHES AND BEST PRACTICES: Data Collection of Audio Dialogues to Support the Training of Speech-to-Speech Translation Systems

by National Institute of Standards and Technology

The purpose of this document is to describe the best practices that personnel from the National Institute of Standards and Technology (NIST) have developed and implemented to efficiently and effectively capture two-way, free-form speech-to-speech audio dialogues within recording studios. These dialogues, produced to support the development and evaluation of machine translation technologies, are conducted by English and foreign language speakers conversing with one another in their native languages through the mediation of an interpreter. NIST personnel have collected over 500 hours of bilingual audio data sets encompassing more than 1100 dialogues across three unique language pairs (English/Iraqi-Arabic, English/Dari, and English/Pashto) since it became involved in this work in 2007. This document will present the methods the NIST team has designed and employed allowing the successful capture of audio data. In addition to the data collection protocols including personnel training and workflow, data collection scenario generation and speaker recruitment protocols will be discussed. NIST Interagency/Internal Report

66 pages, Paperback

First published November 12, 2013