TL;DR
This paper introduces Combilex, a new, high-quality pronunciation lexicon for English, and evaluates its impact on letter-to-sound (LTS) rule generation. By leveraging Combilex's explicit phone-grapheme alignments, the authors demonstrate improved LTS accuracy and significantly accelerated rule development compared to standard automatic alignment methods.
What Problem Does This Paper Address?
Speech synthesizers rely on pronunciation lexicons to convert written text into spoken words. However, existing lexicons are often finite, inconsistent, or limited in scope (e.g., specific accents). This necessitates the development of Letter-to-Sound (LTS) rules to predict pronunciations for words not found in the lexicon. Building accurate LTS rules is challenging due to the inherent complexity and inconsistencies of English spelling. Furthermore, most LTS systems require aligned training data, meaning letters must be explicitly mapped to their corresponding phonetic sounds. This alignment process is often manual and time-consuming, or relies on automatic methods that can introduce errors. This paper addresses the challenge of improving LTS accuracy and development efficiency by introducing and evaluating a novel lexicon, Combilex, which offers explicit, expert-curated phone-grapheme alignments.
Methods
The core of this paper's evaluation lies in comparing LTS models trained using different alignment strategies. The authors use Festival's LTS module, a standard tool for this task.
- Baseline Models (B1, B2): These models were trained using Festival's default "cumulate pairs" method for automatic phone-grapheme alignment. This is a standard approach that attempts to infer alignments from existing data.
- Combilex Models (M1, M2): These models were trained using the explicit, expert-provided phone-grapheme alignments directly from the Combilex lexicon. This is the novel aspect being tested.
The authors also explored two lexicon configurations: a full lexicon and a pruned lexicon (removing words less than four letters long) to see how this affected performance. They also compared their results to a previous study using the Oxford Advanced Learner's Dictionary (OALD).
Datasets
- Combilex Lexicon: A newly developed, high-quality pronunciation lexicon created at CSTR. It features accent-independent "base-form" pronunciations and explicit phone-grapheme alignments. The paper uses a surface-form lexicon derived from Combilex for RP (Received Pronunciation) English, initially containing 143,641 entries, expanded to 212,465 after part-of-speech expansion, and finally filtered to 155,340 words for training.
- Oxford Advanced Learner's Dictionary (OALD): Used for comparison, this lexicon contained approximately 63,399 entries in its original form. The paper references a subset used in a previous study, likely a processed version for LTS training.
Key Insights and Ideas
- Explicit Alignment is Key: The most significant finding is that using expert-aligned phone-grapheme mappings from Combilex leads to demonstrably better LTS rule accuracy than relying on automatic alignment methods. This suggests that the quality of the alignment data is a critical factor in LTS performance.
- Accelerated Development: The explicit alignment provided by Combilex removes the need for manual alignment or complex automatic alignment procedures during LTS model training. This significantly speeds up the development process.
- Combilex Quality: The paper strongly suggests Combilex is a superior lexicon in terms of consistency and size compared to existing options like OALD, which translates to better LTS performance.
- Pruning Short Words: Interestingly, removing short words (less than four letters) from the lexicon improved LTS performance for the Combilex models, but reduced it for the baseline automatic alignment models. This warrants further investigation.
Key Results
- Improved Accuracy: LTS models trained with Combilex's explicit alignment achieved higher accuracy:
- Model M1 (full lexicon): 86.30% words correct, 97.90% phones correct.
- Model M2 (pruned lexicon): 86.50% words correct, 97.99% phones correct.
- This is an improvement over the baseline models (B1: 85.30% words correct, B2: 84.49% words correct).
- Comparison to OALD: Combilex-trained models significantly outperformed models trained on OALD (78.13% words correct).
- Statistical Significance: All improvements in performance between the models were found to be statistically significant at the 1% level.
- Development Speed: Building LTS models is greatly accelerated because the need to list allowed alignments is removed.
Limitations and Future Work
- Test Set Limitations: The authors acknowledge that interpreting LTS test results is inherently problematic. Words that are easily specified in a lexicon are less likely to require LTS prediction, making test sets a potentially poor approximation of real-world use. They also discuss how the presence of multiple word forms (e.g., different parts of speech) or morphologically related words in training/test sets can inflate performance metrics, making it difficult to isolate true generalization ability.
- Future LTS Methods: The authors plan to explore whether other LTS methods, such as pronunciation-by-analogy or joint sequence models, could also benefit from Combilex's expert manual alignment.
- Grapheme-to-Phoneme on Base Forms: They also propose training a grapheme-to-phoneme model directly on Combilex's base-form transcriptions and then converting to surface forms, which would require predicting stress and syllabification.
Summary updated