Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation

Helen M. Meng, Stephanie Seneff, Victor Zue

Abstract

In this paper, we describe a reversible letter-to-sound/sound-to-letter generation system based on an approach which combines a rule-based formalism with data-driven techniques. We adopt a probabilistic parsing strategy to provide a hierarchical lexical analysis of a word, including information such as morphology, stress, syllabification, phonemics and graphemics. Long-distance constraints are propagated by enforcing local constraints throughout the hierarchy. Our training and testing corpora are derived from the high-frequency portion of the Brown Corpus (10,000 words), augmented with markers indicating stress and word morphology. We evaluated our performance based on an unseen test set. The percentage of nonparsable words for letter-to-sound and sound-to-letter generation were 6% and 5% respectively. Of the remaining words our system achieved a word accuracy of 71.8% and a phoneme accuracy of 92.5% for letter-to-sound generation, and a word accuracy of 55.8% and letter accuracy of 89.4% for sound-to-letter generation. We also compared our hierarchical approach with an alternative, single-layer approach to demonstrate how the hierarchy provides a parsimonious description for English orthographic-phonological regularities, while simultaneously attaining competitive generation accuracy.

Introduction: The "New Word" Problem

In the early 1990s, speech synthesis and recognition systems faced a significant challenge known as the "new word" problem. Speech synthesis (Letter-to-Sound or LTS) requires a method to pronounce words not present in its dictionary, while speech recognition (Sound-to-Letter or STL) must guess the spelling of novel terms to expand its vocabulary.

At the time, most solutions were either rigid, hand-crafted rule sets or "black-box" statistical models that lacked linguistic depth. Furthermore, LTS and STL were typically treated as separate problems. This research proposes a unified, reversible framework that handles both tasks by modeling the underlying linguistic logic connecting orthography (spelling) and phonology (sound).

Methods: Hierarchical Lexical Representation

The core of the system is a Hierarchical Lexical Representation. Instead of mapping letters directly to sounds, the system builds a parse tree with seven distinct layers of linguistic structure. This hierarchy allows the system to propagate constraints from high-level morphology down to individual units of sound and spelling.

Hierarchical lexical representation for the word 'dedicate', showing the seven layers of linguistic structure from morphology down to graphemes.

The seven layers include:

  1. Top-level: The word itself.
  2. Morphology: Roots and affixes (e.g., "pre-", "-ing").
  3. Syllable Stress: Identifying stressed vs. unstressed syllables.
  4. Syllable Parts: Onsets, nuclei, and codas.
  5. Broad Classes: Categorizing sounds (e.g., STOP, VOWEL).
  6. Phonemes: Units of sound.
  7. Graphemes: Units of spelling.

The system utilizes a probabilistic parser called TINA and a "layered bigram" formalism. Because the terminal nodes represent both letters and sounds, the system is inherently reversible: to generate sound from spelling, letters are provided as input; to generate spelling from sound, phonemes are used as input.

Datasets and Training

The researchers utilized the Brown Corpus, specifically focusing on the 10,000 most frequent words. The data was split into a training set of approximately 8,000 words and a test set of 800 unseen words. To provide the "ground truth" for the hierarchical layers, words were manually augmented with markers for stress and morphological boundaries.

Key Insights and Efficiency

  • Reversibility via Symmetry: By treating the base of the parse tree as a dual-layer of letters and sounds, the same grammar serves both directions.
  • Hierarchy as a Constraint: High-level knowledge (e.g., syllable stress) constrains lower-level possibilities, making the search space manageable.
  • The "Broad Class" Trap: Interestingly, including "Broad Classes" (Layer 5) slightly hindered performance. This abstraction likely caused "over-smoothing," where the model lost fine-grained distinctions between specific phonemes.
  • Parameter Efficiency: The hierarchical approach is remarkably parsimonious. It achieved performance comparable to the best "single-layer" models of the era while using only 32,000 parameters—a 20-fold reduction compared to the 693,000 parameters required by a flat model.

Key Results

The system was evaluated on "parsable" words (those the grammar could successfully structure). For Letter-to-Sound (LTS) generation, the system achieved 71.8% word accuracy and 92.5% phoneme accuracy on the test set.

Performance results for Letter-to-Sound (LTS) generation, showing word and phoneme accuracy across training and test sets.

For Sound-to-Letter (STL) generation, the system achieved 55.8% word accuracy and 89.4% letter accuracy. The lower word accuracy for STL reflects the inherent ambiguity of English spelling (e.g., the sound /k/ mapping to 'c', 'k', or 'ck'). However, the model's "Top 10" hypotheses for spelling contained the correct answer over 81% of the time, suggesting its utility when paired with a secondary lexicon check.

Limitations and Future Work

Approximately 4-6% of words, primarily foreign names or complex compounds, remained non-parsable by the current grammar. Future work suggests a "robust parsing" mode as a fallback. Additionally, the system struggled with "geminant" letters (double letters like 'tt') that map to a single phoneme. The authors planned to integrate this framework into full speech recognition systems to further explore its bi-directional capabilities.

Summary updated

Abstract

In this paper, we describe a reversible letter-to-sound/sound-to-letter generation system based on an approach which combines a rule-based formalism with data-driven techniques. We adopt a probabilistic parsing strategy to provide a hierarchical lexical analysis of a word, including information such as morphology, stress, syllabification, phonemics and graphemics. Long-distance constraints are propagated by enforcing local constraints throughout the hierarchy. Our training and testing corpora are derived from the high-frequency portion of the Brown Corpus (10,000 words), augmented with markers indicating stress and word morphology. We evaluated our performance based on an unseen test set. The percentage of nonparsable words for letter-to-sound and sound-to-letter generation were 6% and 5% respectively. Of the remaining words our system achieved a word accuracy of 71.8% and a phoneme accuracy of 92.5% for letter-to-sound generation, and a word accuracy of 55.8% and letter accuracy of 89.4% for sound-to-letter generation. We also compared our hierarchical approach with an alternative, single-layer approach to demonstrate how the hierarchy provides a parsimonious description for English orthographic-phonological regularities, while simultaneously attaining competitive generation accuracy.
BibTeX
@inproceedings{meng1994phonological,
  author = {Helen M. Meng and Stephanie Seneff and Victor Zue},
  title = {{Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation}},
  booktitle = {HLT},
  year = {1994},
  url = {https://openreview.net/forum?id=SyE3rm-ObS},
}
Figure