Optimizing Segmentation Strategies For Simultaneous Speech Translation

In this paper, we propose new algorithms for learning segmentation strategies for simultaneous speech translation. In contrast to previously proposed heuristic methods, our method finds a segmentation that directly maximizes the performance of the machine translation system. We describe two methods based on greedy search and dynamic programming that search for the optimal segmentation strategy. An experimental evaluation finds that our algorithm is able to segment the input two to three times more frequently than conventional methods in terms of number of words, while maintaining the same score of automatic evaluation.(1)