Amtrup J.W. Incremental Speech Translation

Springer, 1999. — 212 p.Automatic speech recognition and processing has received a lot of attention during the last decade. Prototypes for speech-to-speech translation are currently being developed that show first impressive results for this highly complex endeavor. They demonstrate that machines can actually be helpful in communicating information between persons speaking different languages. Simple tasks, e.g. the scheduling of business appointments or the reservation of hotel rooms and air travel tickets, are within reach.
Needless to say, the power of these prototypes is far from being equal to the human abilities for speaking, hearing, understanding, and translating. Performing the translation of speeches or free dialog at a high level is one of the most ambitious goals of scientists in the natural language processing domain. Several major areas of research have to be fruitfully combined to create even basic systems and demonstrators. Progress is needed regarding each of the several steps that are performed while creating a translation of an utterance spoken by a speaker, involving fields like acoustics, speech recognition and synthesis, prosody, syntactic processing, semantic representation, contrastive studies for translation, and many others.
This book starts from an outside view to speech translation, a view that does not concentrate immediately on one of the tasks we mentioned. The main motivation for the research presented in this monograph is the fact that humans understand and translate while they are still hearing. This incremental operation is in part responsible for the relative ease with which we handle certain tasks, like simultaneous interpreting or simply following a conversation at a party with a high level of background noise.
The application of this paradigm to automatic speech processing systems seems to be a natural thing to do, yet it has serious consequences for the implementation of individual components and the system as a whole. The simple demand Start analyzing while the input is still incomplete in some cases requires difficult modifications to the algorithms employed for certain tasks.
We think that incremental, modular systems require careful attention as to how they are composed from individual components. Interfaces and their use become more crucial if a component is to deliver not only a small set of final results (in many cases exactly one result), but a continuous stream of hypotheses. Thus, the realization of incrementality also initiated the use of an integrated data structure (the layered chart) and the use of a uniform formalism for all modules.Graph Theory and Natural Language Processing
Unification-Based Formalisms for Translation in Natural Language Processing
MILC: Structure and Implementation
Experiments and Results
Conclusion and Outlook