A learning method for Thai phonetization of English words

Ausdang Thangthai, Chai Wutiwiwatchai, Anocha Ragchatjaroen, Sittipong SaychumInterspeech 2007

TL;DR

This paper tackles the challenge of accurately converting English words into Thai phonetic representations, a crucial step for natural-sounding Thai Text-to-Speech (TTS) systems when encountering English loanwords. They propose a data-driven approach using machine learning (CART) to learn grapheme-to-phoneme (G2P) conversion rules and predict Thai syllabic tones, achieving acceptable results in both objective and subjective evaluations.

What Problem Does This Paper Address?

Modern Thai writing frequently incorporates English words, often written using English orthography within Thai text. When a Thai TTS system encounters these words, simply transcribing them using English phonology results in an unnatural, accented pronunciation. The core problem is to develop a method for accurately converting English words into their Thai phonetic equivalents, respecting Thai phonological rules and syllabic tones, to enable seamless integration into Thai speech synthesis. Existing G2P approaches are often language-specific, and while some research exists for Thai, the unique challenge of integrating English loanwords with Thai phonology, especially regarding syllabic tones, remains a significant hurdle.

Methods

The authors propose a novel framework for English grapheme to Thai phoneme transduction. Their approach is primarily data-driven and utilizes machine learning. Key techniques include:

  • Automatic Grapheme/Phoneme Alignment: Dynamic Time Warping (DTW) is employed to align English graphemes (and phonemes, if available) with corresponding Thai phonemes. A cost function is defined to quantify the matching cost between English and Thai sounds.
  • Grapheme-to-Phoneme Transduction Modeling: Classification and Regression Trees (CART) are used as the learning machine to model the conversion rules. They introduce three variations of the G2P model:
  • EG2TP: A basic model using only English grapheme information.
  • EG2TP-Syl: An enhanced model that incorporates Thai syllable structure constraints by feeding back the preceding predicted Thai phoneme and using a post-processing module.
  • EGP2TP-Syl: The most advanced model, which further leverages English phoneme information (when available from a dictionary) as additional features for the CART model.
  • Syllabic-Tone Prediction: Another CART model is used to predict the appropriate Thai syllabic tone for each syllable. This model utilizes features derived from the syllable structure, vowel/consonant types, and syllable position within the word.

Datasets

  • CMU Pronouncing Dictionary (CMUDICT): Used as a primary source for English word transcriptions. The paper notes its coverage of over 129,000 English words.
  • Custom Dataset: A dataset of 8,300 English words tagged with both CMU and Thai phonological transcriptions was created for training and evaluation. This dataset is crucial as it provides the necessary parallel data for learning the transduction and tone prediction models.

Key Insights and Ideas

  • Leveraging English Pronunciation Dictionaries is Crucial: When an English word is present in a dictionary like CMUDICT, using its English phoneme information significantly improves the accuracy of Thai phonetic transcription compared to relying solely on graphemes. This is because English phonemes can disambiguate grapheme-to-phoneme mappings that are unclear from the spelling alone.
  • Thai Syllable Structure is Key: Incorporating constraints from the Thai syllable structure (initial consonant, vowel, final consonant) into the G2P model (EG2TP-Syl) leads to a substantial improvement in word accuracy, even if phoneme accuracy slightly decreases. This highlights the importance of linguistic structure in phonetic conversion.
  • Syllabic Tone Prediction is Non-Trivial but Achievable with ML: Assigning correct Thai tones to English loanwords is a complex task. The paper demonstrates that a machine learning approach (CART) can effectively learn to predict these tones by considering various syllable-level features.
  • Hybrid Approaches Yield Best Results: The paper advocates for a hybrid approach, combining dictionary lookups with learned rules and constraints, which is a common theme in modern G2P systems.

Key Results

  • G2P Transduction Accuracy:
    • The basic EG2TP model achieved 16.4% word accuracy.
  • EG2TP-Syl (incorporating syllable structure) improved word accuracy to 21.2%.
  • EGP2TP-Syl (using English phonemes) achieved the highest word accuracy at 32.2%.
  • Tone Prediction Accuracy:
    • With ideal inputs (perfect Thai transcriptions), the tone prediction model achieved 71.7% word accuracy.
  • With erroneous inputs (from the G2P model), it achieved 48.1% word accuracy.
  • Integrated System Performance:
    • The integrated EG2TP-Syl + Tone system achieved 16.2% word accuracy.
  • The integrated EGP2TP-Syl + Tone system achieved 24.5% word accuracy.
  • Subjective Evaluation: A listening test with 40 English words showed that over 62% of the synthesized speech was rated as acceptable (Good or Fair).

Limitations and Future Work

  • Ambiguity in Consonant Mapping: The paper acknowledges that some English consonantal characters can be ambiguously mapped to Thai initial consonants, final consonants, or sequences, requiring more training data to resolve.
  • Vowel Length Differentiation: The lack of explicit short/long vowel differentiation in English phonology, while present in Thai, poses a challenge. The determination of Thai vowel length in English words is influenced by factors like lexical stress, syllable position, and word structure, which were not fully incorporated into this model.
  • Future Work Suggestions:
    • Improved Consonant Handling: Develop specific treatments for converting consonants to alleviate ambiguity in distinguishing Thai initial and final consonants.
  • Leveraging Lexical Stress: Utilize lexical stress information from English pronunciation dictionaries to improve vowel length prediction and potentially syllabic tone assignment.

Summary updated

BibTeX
@inproceedings{thangthai2007learning,
  author = {Ausdang Thangthai and Chai Wutiwiwatchai and Anocha Ragchatjaroen and Sittipong Saychum},
  title = {{A learning method for Thai phonetization of English words}},
  booktitle = {Interspeech},
  year = {2007},
}
Figure