Overview of telephone interpretation research at ATR

Abstract

The Automatic Telephone Interpretation system is a facility that enables a person speaking in one language to communicate readily by telephone with someone speaking another. Basic research on the component technologies, speech recognition, machine translation, and speech synthesis is being carried out at ATR. In speech recognition research, the goal is to recognize continuous speech containing a large vocabulary of approximately 3000 words. Based on phonemes as the subunit of speech, words and phrases are recognized. The main approach is a hybrid combination of feature‐based segmentation and phoneme and word recognition by hidden Markov and neural network models. A large‐scale speech database with phonetic transcriptions has been developed for use in speech research. The incorporation of speaker adaptation has been chosen as our approach to speaker independence. Recognition of conversational speech on a specific domain will be explored. Integrating speech and language processing is an important area to be tackled. A technology that uses information about languages and background knowledge to understand the speech content is being studied. As for speech synthesis, rule‐based speech synthesis with high quality and high naturalness is under development. In the area of machine translation, several approaches to the translation of spoken dialogue between Japanese and English are under study. Intention extraction for dialogue interpretation, a plan recognition model for intersentential analysis, and knowledge processing in a specific domain are intended to be investigated.