TTS의 복합 명사 경음 실현 양상에 대한 검토

Abstract

This paper focuses on the phenomenon of tensification as a factor causing differences between the output speech of Text-to-Speech (TTS) systems and human speech. Through TTS output experiments, the realization patterns of tensification within compound nouns were examined among various tensification phenomena. The experimental results revealed a higher incidence of errors in producing compound nouns with tensification when they were supposed to be pronounced with non-tensification, compared to the errors of producing them with tensification when they should be pronounced without it. The error rate in tensification production varied significantly depending on the final consonant type of the preceding morpheme. It is suggested that TTS output errors in tensification within compound nouns result from the program's flawed judgment process, considering only the phonological environment of the preceding and following elements among phonological, morphological, and semantic factors influencing the realization of palatalization. To address these issues, it is proposed to accurately attach the standard pronunciation of vocabulary entries to the database in advance. Additionally, it is emphasized that both morphological and semantic factors should be taken into account in the judgment of tensification within compound nouns, beyond phonological factors.