Sinhala Grapheme-to-Phoneme Conversion and Rules for Schwa Epenthesis

Asanka Wasala, Ruvan Weerasinghe, Kumudu GamageACL 2006

Introduction

Developing accurate Text-to-Speech (TTS) systems for the Sinhala language requires a robust Grapheme-to-Phoneme (G2P) conversion module. This paper, originally presented at COLING/ACL 2006, addresses the "schwa epenthesis" problem—a significant linguistic challenge where the system must determine the correct vowel sound for consonants that lack explicit vowel markers.

The Schwa Epenthesis Problem

Sinhala utilizes an abugida writing system where each consonant character possesses an inherent vowel. When no specific vowel sign is attached, the consonant is pronounced with either a schwa (/ə/) or an /a/ sound. Because the written script does not distinguish between these two, native speakers rely on linguistic intuition. For a computational system, this creates an ambiguity known as schwa epenthesis. Unlike Hindi, which primarily deals with schwa deletion, Sinhala requires a mechanism to insert the correct vowel quality.

Proposed G2P Architecture

The researchers developed a multi-stage rule-based architecture to automate this conversion:

  1. Unicode Mapping and Initial Association: The process begins by mapping Sinhala Unicode characters to base phonemes. By default, ambiguous consonants are associated with the more statistically frequent schwa (/ə/) sound.
Mapping Sinhala graphemes to phonemes using Unicode normalized representations.
  1. Linguistic Rule Application: The core of the system involves eight specific rules designed to refine vowel selection based on syllable position, phonetic context, and word endings. For instance, Rule #1 dictates that a schwa in the initial syllable is typically replaced by /a/.
Application of linguistic rules to refine phoneme sequences.
  1. Diphthong and Exception Processing: The system identifies vowel-semivowel combinations to form diphthongs and consults an exception lexicon for loanwords or irregular forms that do not follow standard phonetic patterns.

Results and Discussion

The system was evaluated against a ground-truth dataset of 30,000 manually transcribed words from the UCSC Sinhala Corpus. The results demonstrate the high efficacy of the rule-based approach:

  • Accuracy: The model achieved a 98% accuracy rate.
  • Impact of Rules: A baseline using only Rule #1 resulted in a 24% error rate. The inclusion of the remaining seven rules reduced the error rate to just 2%, highlighting the importance of detailed linguistic modeling.
  • Error Analysis: The remaining 2% of errors were attributed to compound word boundaries (1.27%), foreign loanwords (0.38%), and homographs that require semantic context for disambiguation.

Conclusion

This research proves that a carefully constructed, linguistically-informed rule-based system can achieve near-perfect G2P conversion for Sinhala. While modern trends often favor deep learning, this approach offers high accuracy with minimal computational overhead. Future improvements will focus on compound word detection and the use of part-of-speech tagging to resolve homograph ambiguities.

Summary updated

BibTeX
@inproceedings{wasala2006sinhala,
  author = {Asanka Wasala and Ruvan Weerasinghe and Kumudu Gamage},
  title = {{Sinhala Grapheme-to-Phoneme Conversion and Rules for Schwa Epenthesis}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2006},
}
Figure