THE ROLE OF PHONETICS AND PHONETIC DATABASES IN JAPANESE SPEECH TECHNOLOGY Jack Halpern (春遍雀來)

This paper summarizes the complex allophonic variations that need to be considered in developing Japanese speech technology applications, with special emphasis on natural speech synthesis, and describes a large-scale phonetic database that provides IPA transcriptions with accent codes and including hundreds of thousands of allophones. The common notion that the kana syllabary correctly represents the Japanese phoneme stock is only partially true. A truly phonemic orthography requires unambiguous grapheme-to-phoneme mappings. Yet speech technology developers often use kana pronunciation dictionaries because of easy availability and low cost. Even more challenging is to accurately generate the allophonic variants and accent patterns for each phoneme and lexeme, which has been done in the project reported in this paper. Index Terms — speech synthesis (TTS), speech recognition (ASR), Japanese allophones, IPA, pitch accen