Unlimited Vocabulary Grapheme to PhonemeConversion with Probabilistic Phrase Break Detection

This paper describes a grapheme-to-phoneme conversion method using phoneme con-nectivity and CCV conversion rules with probabilistic phrase break detection. The method consists of mainly four modules including phrase-break detection, morpheme normalization, morpheme to phoneme conversion and phoneme connectivity check. In the experiments with a test corpus of 210 sentences, we achieved 85% of phrase break detection. The grapheme-to-phoneme conversion performance on the 210 sentences was 85.5% and is improved to 90.8% after employing the phrase break detection. The grapheme-to-phoneme conversion performance on the phrase break free and non-Korean symbol free 4,973 test sentences is 99.9%. The full Korean TTS system is now being implemented using these phrase break detection and grapheme-to-phoneme conversion method.