Statistical Syllable Analysis for Pronunciation Ambiguity Detection and Resolution in Text-to-Speech Synthesis Applications: A Case Study in Turkish

AKBULUT A., Adiguzel T., YILMAZ A. E.

ACTA POLYTECHNICA HUNGARICA, cilt.8, sa.5, ss.53-72, 2011 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 8 Sayı: 5
  • Basım Tarihi: 2011
  • Dergi Adı: ACTA POLYTECHNICA HUNGARICA
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Sayfa Sayıları: ss.53-72
  • Anahtar Kelimeler: Text-to-speech synthesis, natural language processing, grapheme-to-phoneme conversion, less studied languages, pronunciation ambiguity, LEXICON
  • Ankara Üniversitesi Adresli: Evet

Özet

In this study, pronunciation ambiguity in Turkish is considered. A syllable-based ambiguity detection/resolution framework is proposed for Turkish text-to-speech synthesis applications. For this purpose, first the pronunciation ambiguity cases are identified. Such cases are classified into 7 main groups. Statistical analysis on the occurrence rate of these main groups is performed by means of the examination of meaningful Turkish texts. This first level analysis shows that especially the syllables ending with vowels (particularly with a, e and i). which are potential ambiguity sources, have significant occurrence rates. Next, the granularity of the frequency analysis is escalated to distinct syllable level. For the so-far-identified 154 exceptional syllables, the occurrence rates are computed. The results of this study will constitute a major baseline for pronunciation ambiguity detection in Turkish. The resolution of these ambiguous cases will certainly require a large lexicon. The results will also serve as a guideline for the prioritization of data inclusion to such a lexicon (i.e. lexicon enrichment) for rapid coverage. Our distinct syllable level analysis results show that by inclusion of all the words having the 100 most frequent exceptional syllables, it is possible to resolve 99% of pronunciation ambiguities in Turkish. To our belief the findings of this study might also be applicable and useful for other languages.