Extending Pronunciation Lexicons via Non-phonemic Respellings

Lucian GalescuNAACL 2009

TL;DR

This paper introduces a novel approach to extending pronunciation lexicons for speech recognition and text-to-speech systems by leveraging "non-phonemic respellings" (NPRs). These NPRs, which use familiar spelling rules instead of phonetic symbols, make pronunciation specification accessible to average users, significantly improving the accuracy of pronunciation prediction models.

What Problem Does This Paper Address?

Speech recognition (SR) and text-to-speech (TTS) systems rely on pronunciation lexicons (PLs) to map words to acoustic models. Extending these lexicons, whether for specialized domains, new languages, or correcting errors, is a significant challenge. Traditional methods either require expert phonetic knowledge to transcribe words into complex phoneme sets or use limited "pronunciation-by-orthography" methods. Automatic grapheme-to-phoneme (G2P) converters, while helpful, often fail when users need to provide their own pronunciations, as they require users to understand phonetic alphabets if the system's suggestions are incorrect. This paper aims to bridge this gap by enabling average users to contribute pronunciation information easily.

Methods

The core of the paper's contribution lies in the use of Non-Phonemic Respellings (NPRs). NPRs represent pronunciations using familiar English spelling rules, often with capitalization or other markers to indicate stress (e.g., \FO-neem\ for "phoneme"). This approach bypasses the need for users to learn phonetic alphabets.

The authors then investigate how to convert these NPRs into actual phoneme pronunciations. They explore three main strategies:

  1. Spelling-based G2P: A standard grapheme-to-phoneme model trained on the CMU dictionary.
  2. NPR-based N2P: A novel "non-phonemic respelling-to-phoneme" (N2P) model trained on the NPRs themselves.
  3. Combined Model: An oracle combination of the spelling-based G2P and the NPR-based N2P models.

They also address the challenge of adapting generic G2P models to the N2P task by identifying specific grapheme-phoneme correspondences that are common in NPRs but rare in standard dictionaries. They propose post-processing rules to handle these discrepancies and suggest that a more adequate N2P model could be built by selecting relevant entries from generic dictionary data.

Datasets

  • National Cancer Institute's Dictionary of Cancer Terms: A corpus of 2730 words with a total of 2847 NPR transcriptions. This dataset was split into:
  • ID set: 1183 words also present in the CMU dictionary.
  • OOD set: 928 words not found in the CMU dictionary.
  • CMU Pronouncing Dictionary (CMUdict): Used for training the baseline G2P model and for identifying entries with NPR-specific grapheme-phoneme mappings.
  • Merriam-Webster Online Medical Dictionary and New Oxford American Dictionary: Used to verify phonetic transcriptions for the OOD set.
  • User-generated NPRs: A small dataset collected from five computer-savvy subjects for evaluating user-generated NPRs.

Key Insights and Ideas

  • NPRs as an Accessible Interface: The fundamental idea is that NPRs provide a user-friendly way for non-experts to specify pronunciations, democratizing lexicon extension.
  • NPRs Capture Different Phonological Rules: NPRs often employ grapheme-phoneme correspondences that are rare or absent in standard dictionaries, necessitating specialized N2P models.
  • Generic G2P Models Need Adaptation: While generic G2P models can be a starting point, they are not directly sufficient for N2P tasks due to the unique nature of NPRs. Targeted adaptation and rule-based corrections are crucial.
  • Combined Models Offer Significant Gains: Integrating NPR information with standard spelling-based G2P models leads to substantial improvements in pronunciation prediction accuracy, especially for out-of-vocabulary words.
  • Novice Users Can Generate Useful NPRs: Preliminary studies suggest that even users with limited exposure to NPRs can produce reasonably accurate respellings, indicating the approach's practical viability.

Key Results

  • Improved Accuracy on ID Set: The combined spelling and NPR-based model achieved significant error rate reductions compared to the G2P model alone: over 57% reduction in Phone Error Rate (PER) and over 47% reduction in Word Error Rate (WER) for the top hypothesis.
  • Improved Accuracy on OOD Set: For out-of-vocabulary words, the N2P model showed almost 19% better PER and 5% better WER than the G2P model. The combined model achieved 49% better PER and 31% better WER.
  • User-Generated NPR Performance: For user-generated NPRs, the G2P model had an average PER of about 45%. NPRs used with the G2P model (without new GP mappings) reduced PER to around 36% (+/-5%), and the combined model achieved about 33% (+/-5%) PER.

Limitations and Future Work

  • Preliminary Nature: The authors acknowledge that the experiments are preliminary and more data collection and analysis are needed.
  • NPR Variation: Users can generate diverse NPRs, sometimes deviating from standard rules. The paper suggests an error model might be needed to handle such issues.
  • Limited User Study: The user study was small, and further investigation is needed to assess average users' ability to generate NPRs and how their respellings deviate from general English rules.
  • Future Work:
    • Collecting and analyzing more data for training robust NPR-to-pronunciation models.
  • Investigating the reverse problem: obtaining NPRs from pronunciations.
  • Developing systems that can provide feedback, including generating NPRs.
  • Further exploring how novice users' NPR generation skills evolve with practice.

Summary updated

BibTeX
@inproceedings{galescu2009extending,
  author = {Lucian Galescu},
  title = {{Extending Pronunciation Lexicons via Non-phonemic Respellings}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
}
Figure