International Workshop on Spoken Language Translation (IWSLT) 2006
Keihanna Science City, Kyoto, Japan
5 years of machine translation evaluations, esp. those organized by NIST,
tell a wonderful success story: machine translation has made amazing progress,
at least when translating text sources, like news, for the "big" languages
like Arabic and Chinese. This improvement comes from using more data in
data-driven MT systems, but also from better models, better training and
The picture seems less sparkling when we look at speech translation. Even in a simple approach, in which first-best recognition result is fed into a text translation system the improvements in speech recognition and in machine translation will result in improved speech translation. There has been quite some work on tighter coupling between the two components, like lattice translation, using ASR features in MT and optimizing the end-to-end performance, it still seems to be once-in-a-while improvements which result from this work. Additional topics relating to speech translation are disfluency removal, segmentation and automatic punctuation.
The panel discussion will take a critical look at recent work in speech translation:
Bibliographic reference. Vogel, Stephan / Dillinger, Mike / Federico, Marcello / Zhang, Ruiqiang (2006): "Speech translation - where is the progress? what are the challenges?", In IWSLT-2006 (panel).