Finite-State Acoustic and Translation Model Composition in Statistical Speech Translation: Empirical Assessment

Speech translation can be tackled by means of the so-called decoupled approach: a speech recognition system followed by a text translation system. The major drawback of this two-pass decoding approach lies in the fact that the translation system has to cope with the errors derived from the speech recognition system. There is hardly any cooperation between the acoustic and the translation knowledge sources. There is a line of research focusing on alternatives to implement speech translation efficiently: ranging from semi-decoupled to tightly integrated approaches. The goal of integration is to make acoustic and translation models cooperate in the underlying decision problem. That is, the translation is built by virtue of the joint action of both models. As a side-advantage of the integrated approaches, the translation is obtained in a single-pass decoding strategy. The aim of this paper is to assess the quality of the hypotheses explored within different speech translation approaches. Evidence of the performance is given through experimental results on a limited-domain task.