Incremental Speech Translation

Human language capabilities are based on mental proceduresthat are closely linked to the time domain. Listening, understanding,and reacting, on the one hand, as well as planning,formulating,and speaking,onthe other, are performedin a highlyover lapping manner, thus allowing inter human communication to proceed in a smooth and ?uent way. Although it happens to be the natural mode of human language interaction, in cremental processing is still far from becoming a common feature of today’s lan guage technology. Instead, it will certainly remain one of the big challenges for research activities in the years to come. Usually considered dif?cult to a degree that rendersit almost intractableforpracticalpurposes,incrementallanguageprocessing has recently been attracting a steadily growing interest in the spoken language pro cessing community. Its notorious dif?culty can be attributed mainly to two reasons: Due to the inaccessibility of the right context, global optimization criteria are no longer available. This loss must be compensated for by communicating larger search spaces between system components or by introducing appropriate repair mechanisms. In any case, the complexity of the task can easily grow by an order of magnitude or even more. Incrementality is an almost useless feature as long as it remains a local property of individual system components. The advantages of incremental processing can be effectiveonly if all the componentsof a producer consumerchain consistently adhere to the same pattern of temporal behavior.

In 1992 it seemed very difficult to answer the question whether it would be possible to develop a portable system for the automatic recognition and translation of spon taneous speech. Previous research work on speech processing had focused on read speech only and international projects aimed at automated text translation had just been terminated without achieving their objectives. Within this context, the German Federal Ministry of Education and Research (BMBF) made a careful analysis of all national and international research projects conducted in the field of speech and language technology before deciding to launch an eight-year basic-research lead project in which research groups were to cooperate in an interdisciplinary and international effort covering the disciplines of computer science, computational linguistics, translation science, signal processing, communi cation science and artificial intelligence. At some point, the project comprised up to 135 work packages with up to 33 research groups working on these packages. The project was controlled by means of a network plan. Every two years the project sit uation was assessed and the project goals were updated. An international scientific advisory board provided advice for BMBF. A new scientific approach was chosen for this project: coping with the com plexity of spontaneous speech with all its pertinent phenomena such as ambiguities, self-corrections, hesitations and disfluencies took precedence over the intended lex icon size. Another important aspect was that prosodic information was exploited at all processing stages.

The 12th Australian Joint Conference on Artificial Intelligence (AI'QQ) held in Sydney, Australia, 6-10 December 1999, is the latest in a series of annual re gional meetings at which advances in artificial intelligence are reported. This series now attracts many international papers, and indeed the constitution of the program committee reflects this geographical diversity. Besides the usual tutorials and workshops, this year the conference included a companion sympo sium at which papers on industrial appUcations were presented. The symposium papers have been published in a separate volume edited by Eric Tsui. Ar99 is organized by the University of New South Wales, and sponsored by the Aus tralian Computer Society, the Commonwealth Scientific and Industrial Research Organisation (CSIRO), Computer Sciences Corporation, the KRRU group at Griffith University, the Australian Artificial Intelligence Institute, and Neuron- Works Ltd. Ar99 received over 120 conference paper submissions, of which about o- third were from outside Australia. Prom these, 39 were accepted for regular presentation, and a further 15 for poster display. These proceedings contain the full regular papers and extended summaries of the poster papers. All papers were refereed, mostly by two or three reviewers selected by members of the program committee, and a list of these reviewers appears later. The technical program comprised two days of workshops and tutorials, fol lowed by three days of conference and symposium plenary and paper sessions.

This book presents a collection of state-of-the-art work in corpus-based interpreting studies, highlighting international research on the properties of interpreted speech, based on naturalistic interpreting data. Interpreting research has long been hampered by the lack of naturalistic data that would allow researchers to make empirically valid generalizations about interpreting. The researchers who present their work here have played a pioneering role in the compilation of interpreting data and in the exploitation of that data. The collection focuses on both of these aspects, including a detailed overview of interpreting corpora, a collective paper on the way forward in corpus compilation and several studies on interpreted speech in diverse language pairs and interpreter-mediated settings, based on existing corpora.

The series serves to propagate investigations into language usage, especially with respect to computational support. This includes all forms of text handling activity, not only interlingual translations, but also conversions carried out in response to different communicative tasks. Among the major topics are problems of text transfer and the interplay between human and machine activities.

Proceedings of the NATO Advanced Study Institute on Computational Models of Speech Pattern Processing, held in St. Helier, Jersey, UK, July 7-18, 1997

Automatic Speech Translation introduces recent results of Japanese research and development in speech translation and speech recognition. Topics covered include: fundamental concepts of speech recognition; speech pattern representation; phoneme-based HMM phoneme recognition; continuous speech recognition; speaker adaptation; speaker-independent speech recognition; utterance analysis, utterance transfer, utterance generation; contextual process­ing; speech synthesis and an experimental system of speech translation. This book presents the complicated technological aspects of machine translation and speech recognition, and outlines the future directions of this rapidly developing area of technology.