Stochastic pronunciation modelling from hand-labelled phonetic corpora

@article{riley1999pronunciation,
title = {Stochastic pronunciation modelling from hand-labelled phonetic corpora},
journal = {Speech Communication},
volume = {29},
number = {2},
pages = {209-224},
year = {1999},
issn = {0167-6393},
doi = {https://doi.org/10.1016/S0167-6393(99)00037-0},
url = {https://www.sciencedirect.com/science/article/pii/S0167639399000370},
author = {Michael Riley and William Byrne and Michael Finke and Sanjeev Khudanpur and Andrej Ljolje and John McDonough and Harriet Nock and Murat Saraclar and Charles Wooters and George Zavaliagkos},
}

📑 Outline of Riley et al. (1999)

1. Introduction

  • Motivation for modeling pronunciation variation
  • Challenges of canonical dictionaries in ASR
  • Advantages of learning from phonetically transcribed speech corpora

2. Pronunciation Modeling Approach

  • Use of decision trees to learn mappings between:

    • Lexical (canonical) phoneme sequences
    • Observed surface phone sequences

  • Model representation:

    • Input: phoneme with context (e.g., word position, neighboring phonemes)
    • Output: probabilistic distribution over phones

  • Description of how variants are extracted from hand-labeled data

3. Corpora Used

  • Description of three phonetic corpora:

    1. TIMIT: Read speech, clean recordings
    2. ICSI Switchboard corpus: Conversational telephone speech
    3. CSLU Spontaneous Speech corpus

  • Details of:

    • Lexicon preparation
    • Forced alignment and phonetic transcription alignment methods

4. Data Preparation

  • Canonical pronunciation lexicon derivation
  • Aligning dictionary forms with hand-labeled surface realizations
  • Mapping phones to phonemes (many-to-one and one-to-many mappings)

5. Decision Tree Training

  • Learning pronunciation rules using CART-style decision trees

  • Feature set for conditioning:

    • Left/right phonemic context
    • Word position
    • Stress markers
    • Part-of-speech tags (optional)
  • Smoothing and pruning strategies

6. Experimental Results

  • Evaluation metrics:

    • Word error rate (WER) improvement in ASR
    • Lexicon coverage increase
    • Comparison of baseline vs. expanded pronunciation lexicons

  • Results on:

    • TIMIT (limited gain, due to low variation)
    • ICSI (more natural variation, larger gains)
    • CSLU (high variability, challenging data)

7. Analysis and Discussion

  • Interpretation of decision tree outputs

  • Nature of learned pronunciation rules:

    • Context-dependent deletions, insertions, substitutions

  • Examples of learned patterns:

    • Flapping, vowel reduction, glottalization
  • Observations on overfitting and rule generalization

8. Comparison with Other Approaches

  • Rule-based vs. data-driven methods
  • Relation to maximum entropy models
  • Contrast with G2P models trained only on dictionaries

9. Conclusion

  • Summary of findings:

    • Data-driven pronunciation modeling from hand-labeled corpora is effective
    • Learns interpretable and generalizable variation

  • Limitations:

    • Dependency on high-quality phonetic corpora

  • Future work:

    • Integration with G2P systems
    • Use of more diverse corpora
    • Better handling of spontaneous speech variation