Spontaneous dialogue recognition with large out-of-vocabularies

Abstract

This paper describes a Korean spontaneous speech recognition system that is part of a speech-to-speech translation system being developed in ETRI. The features of this system are as follows. First, we have compiled two large human to human dialog databases. The speech database used in training and evaluation of this system was composed of two task domains: 339 dialogs collected in appointment scheduling domain and 535 dialogs collected in the travel planning domain. Second, recognition performance was evaluated not only on no out-of-vocabulary (OoV) case but also on the OoV case. From the evaluation result of spontaneous speech recognition with high OoV rate, we addressed the problem of Korean language and proposed a solution.