Rapid Development Techniques for Spoken Language Translation

Current approaches to building speech translation systems require large amounts of data for building acoustic, language, and translation models. The data requirements lead to high costs and long lead times when building systems for new languages or domains. Sehda, Inc. proposes to dramatically reduce the time and cost to create two-way speech translation systems in new languages and domains by creating new algorithms and approaches to system development that: • Reduce the amount of training data with a divide-and-conquer SMT approach . • Augment existing domain corpora with similar data from the Internet in both English and the target language. • Use acoustic models from other languages to create models in a new language. • Build better translation models from the limited domain-specific parallel corpus. The goal of Phase I is to define a rapid system-building process using these innovations and to deliver a working prototype for a single domain in Korean. In Phase II, Sehda will create a system in another language that is structurally different from Korean and create new algorithms needed for that language, expanding the number of tools that are available to build systems for other languages.