Thai Grapheme-Based Speech Recognition

Paisarn Charoenpornsawat, Sanjika Hewavitharana, Tanja SchultzNAACL 2006

TL;DR

This paper explores grapheme-based speech recognition (GBSR) for Thai, a language with complex orthography. They demonstrate that GBSR, particularly when enhanced with a novel tree clustering technique, can achieve performance comparable to traditional phoneme-based systems, even those using hand-crafted dictionaries, thereby reducing the need for extensive linguistic expertise.

What Problem Does This Paper Address?

Traditional large-vocabulary speech recognition systems rely on phonemes as sub-word units. This necessitates the creation of a pronunciation dictionary that maps written words to their phonetic representations. For many languages, especially those with less standardized orthography or complex writing systems, generating these dictionaries is a time-consuming and labor-intensive process requiring significant linguistic knowledge. Grapheme-based speech recognition (GBSR) offers a potential solution by using graphemes (written characters or character sequences) as the sub-word units, simplifying dictionary creation. However, the looser correspondence between graphemes and pronunciation compared to phonemes presents challenges, requiring more sophisticated modeling techniques to capture pronunciation variations. This paper investigates the effectiveness of GBSR for Thai, a language with unique orthographic features, and explores methods to improve its performance.

Methods

The authors experiment with several approaches to build their Thai GBSR system:

  • Grapheme-to-Phoneme Mapping: They address Thai's complex writing system by applying preprocessing steps to better align graphemes with phonemes. This includes rules for reordering graphemes when vowels precede consonants and for merging multiple letters representing a single phoneme. Initially, they found reordering alone yielded better results.
  • Bootstrapping Acoustic Models: To overcome the initial "flat-start" approach (where acoustic models are initialized with uniform distributions), they "bootstrap" their grapheme-based models from pre-trained phoneme-based systems. They compare bootstrapping from a Thai-specific phoneme-based system versus a multilingual one, finding the latter to be more effective for their experiments.
  • Context-Dependent Modeling: They explore using different grapheme units for context-dependent modeling, specifically trigraphemes (a grapheme and its immediate neighbors) and quintgraphemes (a grapheme and two neighbors on each side). They also investigate the impact of acoustic model size (500 vs. 2000 models).
  • Enhanced Tree Clustering (ETC): A key contribution is the application and evaluation of an enhanced tree clustering method. Unlike traditional decision trees that build separate trees for each phoneme sub-state, ETC builds trees that cover multiple phonemes. This allows for greater sharing of acoustic model parameters across similar-sounding phonetic contexts, aiming to improve robustness and reduce data sparsity issues inherent in grapheme-based systems.

Datasets

  • Thai GlobalPhone Corpus: This corpus, consisting of 20 hours of read speech from 90 native Thai speakers, was used for training. It covers approximately 260,000 words with a vocabulary of about 7,400 words.
  • Test Set: 1,181 utterances from 8 different speakers were used for testing.
  • Language Model: A language model was built on news articles, resulting in a trigram perplexity of 140 and an Out-Of-Vocabulary (OOV) rate of 1.4% on the test set.

Key Insights and Ideas

  • GBSR Viability for Thai: The paper demonstrates that grapheme-based speech recognition is a viable approach for Thai, despite its complex orthography.
  • Bootstrapping is Crucial: A simple "flat-start" initialization for acoustic models leads to poor performance. Bootstrapping from existing phoneme-based models significantly improves results.
  • Enhanced Tree Clustering's Power: The enhanced tree clustering (ETC) method provides a substantial performance boost for the grapheme-based system. This highlights the importance of effective parameter sharing for handling pronunciation variations in grapheme-based systems.
  • Trigraphemes Outperform Quintgraphemes: For context-dependent units, trigraphemes yielded better results than quintgraphemes, likely due to data sparseness issues with longer units.
  • Reducing Linguistic Burden: The success of GBSR, especially with ETC, suggests a pathway to developing speech recognition systems for new languages without the extensive manual effort typically required for phonetic dictionary creation.

Key Results

  • The best grapheme-based system with trigraphemes and 500 acoustic models achieved a Word Error Rate (WER) of 26.0%.
  • When enhanced tree clustering (ETC-500) was applied to this system, the WER dropped significantly to 16.9%.
  • This ETC-500 system achieved comparable results to a phoneme-based system using a hand-crafted dictionary (16.0% WER).
  • The grapheme-based system with ETC-500 significantly outperformed a phoneme-based system that used an automatically generated dictionary (24.5% WER).

Limitations and Future Work

  • Data Sparseness: The paper notes that quintgraphemes performed worse than trigraphemes, suggesting data sparseness is a challenge when using longer context units.
  • Acoustic Model Size: While increasing acoustic models from 500 to 2000 did not improve trigrapheme performance, it did slightly worsen quintgrapheme performance. The optimal model size might depend on the context unit and modeling technique.
  • ETC Parameter Tuning: The paper observed that ETC-500 (500 Gaussians per codebook) outperformed ETC-1500 (1500 Gaussians), suggesting that increasing the number of leaf nodes in the cluster trees (which is related to the number of Gaussians) can sometimes degrade performance. Further tuning of ETC parameters could be beneficial.
  • Future Work: The authors suggest that the promising results indicate that GBSR can be successfully applied to new languages with minimal linguistic knowledge. Further research could explore more advanced grapheme-to-phoneme mapping techniques and more sophisticated acoustic modeling strategies for Thai.

Summary updated

BibTeX
@inproceedings{charoenpornsawat2006thai,
  author = {Paisarn Charoenpornsawat and Sanjika Hewavitharana and Tanja Schultz},
  title = {{Thai Grapheme-Based Speech Recognition}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2006},
}
Figure