Chapter 10 - Speech-to-Speech Translation

Abstract

This chapter addresses the theory and practice of speech-to-speech translation, including different translation approaches and strategies for coupling speech recognition and translation. Speech-to-speech translation is the task of translating speech input in one language into speech output in another language. This process consists of three basic steps: speech recognition, translation, and speech generation. There are various approaches to speech-to-speech translation, including interlingua-based, example-based, statistical, and transfer approaches. Speech translation consists of three sources of errors: errors introduced by the speaker, errors of the recognizer, and errors of the translation system. The first part of this chapter compares interlingua-based and statistical implementations within the framework of the NESPOLE! system developed at Carnegie Mellon University. It discusses decoding strategies allowing for flexible phrase reordering and spoken language-specific problems. The second part focuses on example-based and transfer approaches and describes their realization within the spoken language translation system developed at advanced telecommunications research institute international (ATR) in Japan.