Vowel Sound Disambiguation for Intelligible Korean Speech Synthesis

Proceedings of PACLIC 19, the 19th Asia-Pacific Conference on Language, Information and Computation. Vowel Sound Disambiguation for Intelligible Korean Speech Synthesis Ho-Joon Lee Computer Science Division EECS department, KAIST 373-1 Guseong-dong, Yuseong-gu Daejeon 305-701 South Korea hojoon@nlp.kaist.ac.kr Jong C. Park Computer Science Division EECS department, KAIST 373-1 Guseong-dong, Yuseong-gu Daejeon 305-701 South Korea park@nlp.kaist.ac.kr Abstract For speech synthesis systems that transform text materials into voice data, correctness and naturalness are the crucial measures of performance, the latter gaining more emphasis recently. In order to make synthesized voices natural, we must take into account pronunciation-related linguistic phenomena such as homograph, among others. The syntax certainly provides an important clue to disambiguating such homographs, but the relatively free word order in the Korean language makes it hard to utilize such information. In this paper, we describe a computational generation of contextually appropriate vowel lengths for the words in Korean by utilizing a higher level of linguistic information in a Combinatory Categorial Grammar framework. We consider parts-of- speech information, the possibility of conjunction with a suffix, case information, unconjugated adjectives, numerals, numerical adjectives with related nouns, and the relationship between a noun and its predicate as syntactic and semantic clues for vowel sound disambiguation. The results are expressed in Speech Synthesis Markup Language (SSML) for a target system neutral application. The proposed system with correctly predicted vowel sound can be used not only as an educational tool, but also as a plug-in for enhancing the intelligibility of a general purpose Text-to-Speech (TTS) system. 1. Introduction Speech plays an important role as a communication method from birth to death, and it has also been regarded as the most user-friendly interface to machines thanks to the great development of the Information Technology such as the technologies for computers and mobile phones. For speech synthesis systems that transform text sources into voice data, correctness and naturalness are considered the crucial measures of performance. The naturalness enhances usability of speech synthesis systems with adding intelligibility to synthesized results. In order to make synthesized voices more natural, we must take into account pronunciation-related linguistic phenomena, such as homograph, among others. While most homographs in English have different vowel sound such as ‘read [ri:d]’ versus ‘read [red]’ and ‘bass [beis]’ versus ‘bass [bæs]’, homographs in Korean share nearly the same pronunciation with one another, except for the lengths of their vowel sound. This has apparently caused many people even including television news announcers, to suffer from the confusion between short vowel sound and long vowel sound in homographs. Disambiguation of short and long vowel sound for a homographic word in Korean represents not only the standard pronunciation, but also a mark of different meanings of that word. For instance, the word ‘밤(pam)’ has short vowel sound ‘[bam]’ to mean ‘night’, and long vowel sound ‘[ba:m]’ to mean ‘chestnut’. The following examples also show two different meaning of the word ‘눈(nwun)’ that means ‘eyes’ with short vowel sound ‘[nun]’, and ‘snow’ with long vowel sound ‘[nu:n]’, respectively. (1) 철수가 눈이 …

Content truncated to 500 words