Multilingual Speech Interpretation for Cellular Phones
Yasunari Obuchi, Yoshinori Kitahara, 敦子 小泉 · 2001
This paper describes the development and field trial tests of the multilingual speech interpretation system. The system is based on the fixed sentence interpretation, in order to achieve a high recognition rate and to guarantee the translation accuracy. More than 1,700 sentences are registered, and a simple network grammar enables the system to deal with various expression forms for the same meaning. The first field trial was carried out with the Japanese-to-Korean interpretation system. We received 6,753 calls during two months, and the data showed that the biggest problem was the out-of-vocabulary (OOV) inputs, which resulted in 39% task completion rate. The second prototype system, that has the output of ten languages, was improved in view of the result of the first field trial. We added various expression forms found in the first trial data, and we also introduced the OOV detecting module using garbage models. The second field trial results showed that those improvements reduced misrecognitions due to the OOV inputs. The paper also describes the prototype of the bi-directional interpretation system between Japanese and Korean. We adapted the Hidden Markov Models (HMMs) made by the Japanese speech data to the Korean speech recognition system. The experimental results showed that the system works well for the fixed sentence recognition, even though its robustness in other tasks and different conditions remains to be discussed.