Disambiguation of Korean Utterances Using Automatic Intonation Recognition

Abstract

The paper describes a research on a use of intonation for disambiguating utterance types of Korean spoken sentences. Based on tilt intonation theory (Taylor and Black 1994), two related but separate experiments were performed at speaker independent level, both using the Hidden Markov Model training technique. In the first experiment, a system is established so that rough boundary positions of major intonation events are detected. Subsequently the significant parameters are extracted from the products of the first experiment, which are directly used to train the final models for utterance type disambiguation. Results show that the intonation contour can be used as a significant meaning distinguisher in an automatic speech recognition system of Korean as well as in a natural human communication system.