Abstract
Because of recognition errors, the performance (quality) of speech translation is degraded. Previously, we proposed a method that used only reliable parts of the recognition result for the translation. However, in this method, non-translated parts are omitted even if useful information exist in these parts. To overcome this problem, we propose an error correction method which is composed of the following steps: (1) The necessity of correction is judged and only utterances of the recognition results with “potentially” recoverable erroneous parts are retained.(2) The example utterances that have phonetically similar parts to the ones retained in the step (1) are retrieved from a text corpus, and correction hypotheses are created.(3) The reliability of the correction hypotheses is judged according to both semantic and phonetic point of view and the most reliable one is selected. The error correction method was incorporated into a speech translation system, and evaluated for speech inputs in travel conversations. As the results, the word error rate was reduced by 2.3%, and the acceptable translations rate was increased by 5.4%.
