The BBN TransTalk Speech-to-Speech Translation System

Portable translation devices which enable people who speak different languages to communicate with each another in voice are likely to have a far-reaching impact in both the civilian and military worlds. While long a staple of science fiction, a la the "Star Trek universal translator", devices that actually translate between languages have not been fully realized. In the last decade, however, under the sponsorship of the Babylon and TRANSTAC (Translation for Tactical use) programs of the US Defense Advance Research Program Agency (DARPA), several research sites, including BBN (Stallard et al., 2007), CMU (Waibel et al., 2003), IBM (Gao et al., 2006), SRI (Akbacak et al., 2009) and USC (Belvin et al. 2005), have made significant progress in developing a real two-way speech-to-speech (S2S) translation systems. These systems are not “universal translators” in the sciencefiction sense, in that must be configured for the language and conversational domain of interest, rather than spontaneously understanding them. However, the technology is language-independent, and under the auspices of the TRANSTAC program, systems have been configured for several different foreign languages of interest to the US Government, including Iraqi Arabic, Malay, Farsi, Dari, and Pashto. Though the technology is also domain-independent, most of these systems support conversations in the so-called "force protection" military domain, which is broadly construed to include not only conversations relevant to checkpoints, searches, and other military operations, but also rapport building, civil affairs, and basic medical conversations. In this article, we describe BBN’s S2S system, TransTalk, which runs not only on laptops and ultra-mobile PCs, but also on mobile Android Smartphones, running locally on the device itself and not a server. In common with other TRANTSTAC systems, TransTalk’s technology is language-independent, and has been configured to translate between English and numerous other languages, including Iraqi Arabic, Dari, Pashto, Farsi, and Malay. TransTalk also has a uniquely simple user interface which does not require the user to view a screen. TransTalk integrates automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS) synthesis engines for converting speech in one language into a different language. In particular, TransTalk uses the BBN Byblos ASR engine for converting speech to text. BBN Byblos is a multi-pass, speaker-independent large