TL;DR
This paper tackles the challenge of open-vocabulary Thai speech recognition by proposing a novel hybrid language model that combines full words with "pseudo-morphemes" (syllable-like units). This approach significantly reduces out-of-vocabulary errors and achieves comparable recognition accuracy with a much smaller lexicon, making it more practical for real-world applications.
What Problem Does This Paper Address?
Developing large-vocabulary continuous speech recognition (LVCSR) systems for languages like Thai presents a significant hurdle: the "open-vocabulary" problem. This means the system must handle an ever-growing lexicon, including new words, names, and technical terms that weren't present during training. Thai, in particular, exacerbates this issue because it lacks word boundary markers, making word segmentation ambiguous, and it frequently uses compound words. Traditional LVCSR systems struggle with out-of-vocabulary (OOV) words, leading to recognition errors. While sub-word units (like morphemes) are a common solution for OOV words in other languages, Thai's linguistic structure makes morphemes unsuitable.
Methods
The core innovation here is a hybrid language model that combines two types of units:
- Full Words: Frequently occurring words are directly included in the lexicon.
- Pseudo-Morphemes (PMs): These are syllable-like units derived from Thai text. Unlike linguistic morphemes, PMs are based on the written form and are more deterministic for Thai. The idea is that by combining these smaller, more predictable units, new or rare words can be constructed and recognized.
The authors developed a process to create this hybrid lexicon:
- Identify Frequent Words: Words appearing above a certain threshold in the training data are kept as full words.
- Segment Rare Words into PMs: Words not meeting the frequency threshold are segmented into PMs using a syllable segmentation tool that leverages Thai writing rules.
- Combine Lexicons: The final lexicon comprises frequent full words and the PMs derived from less frequent words.
- Train Language Model: A trigram language model is trained using this hybrid lexicon, employing a modified Kneser-Ney discounting technique.
Datasets
The research utilized three large Thai text and speech corpora for training and language model development:
- BEST: A large Thai text corpus (approx. 7 million words) manually segmented into words, covering academic articles, encyclopedias, novels, and news.
- LOTUS-BN: A Thai broadcast news corpus with audio recordings and transcriptions (approx. 100 hours of speech from 81 speakers). This was used for acoustic model training and language model training (specifically the training set, TR).
- HIT-BTEC: A parallel Thai-English corpus, originally developed for speech-to-speech translation research, containing about 160,000 utterances.
For evaluation, data from three different recognition tasks were used:
- Broadcast News Transcription (BN): 2,200 utterances from the LOTUS-BN evaluation test set.
- Dictation: 300 utterances recorded in an office environment, covering various genres.
- Mobile Speech-to-Speech Translation (S2S): 2,000 utterances recorded on mobile devices in sport and travel domains.
Key Insights and Ideas
- PMs as a Viable Sub-lexical Unit for Thai: The paper successfully demonstrates that syllable-like units (PMs), derived from the written form, are a suitable alternative to linguistic morphemes for handling OOV words in Thai LVCSR.
- Hybrid Model's Effectiveness: Combining full words with PMs is crucial. Relying solely on PMs leads to acoustic confusability and shorter context spans in language models. The hybrid approach balances these drawbacks.
- Lexicon Size Reduction: The hybrid model allows for a significantly smaller lexicon size (e.g., 35K entries) while still achieving good performance, making it more practical for resource-constrained environments like mobile devices.
- OOV Reduction is Key: The primary benefit of the hybrid model is its dramatic reduction in OOV rates, often to below 1%, even outperforming a full-word lexicon that uses a much larger vocabulary.
Key Results
- OOV Rate Reduction: The hybrid LMs consistently reduced OOV rates to less than 1% across all test sets, a significant improvement over baseline full-word LMs (which had OOV rates around 1.65% to 3.24%).
- Recognition Error Rate (PER): The best hybrid LM (H1, using 20K words and 15K PMs) achieved the best PER on the BN and Dictation test sets, reducing it by 0.50% absolute (1.37% relative) and 0.56% absolute (1.70% relative) respectively, compared to a full-word baseline with a similar lexicon size.
- Comparable Performance on S2S: On the S2S test set, the hybrid LM's recognition rate was comparable to the baseline.
- No Negative Impact on In-Vocabulary Words: The study found that including PMs did not negatively affect the recognition of in-vocabulary words (MIV - misrecognized in-vocabulary words - was not higher than the baseline).
- Perplexity (PP): While the hybrid model generally improved performance, perplexity increased significantly when too many PMs were introduced (e.g., in H3), indicating potential issues with context span.
Limitations and Future Work
- Acoustic Confusability: The paper notes that while PMs help with OOV, introducing too many small units can increase acoustic confusability.
- Short Context Span: Language models built on small units like PMs can suffer from shorter context spans, leading to increased perplexity.
- Pronunciation Variations: The authors suggest that to further improve recognition accuracy, the pronunciation variations of PMs should be considered.
- Variable-Length N-gram Models: To address the short context span issue, they propose exploring variable-length n-gram language models.
Summary updated