TL;DR
This paper explores how the frequency and context in which English phonemes appear influence how native speakers perceive their similarity. By building a "distributional similarity" model, the authors show it better captures human perception than traditional feature-based models, suggesting that understanding phoneme co-occurrence is crucial for modeling sound perception.
What Problem Does This Paper Address?
Understanding how speakers perceive the similarity between sounds (phonemes) is fundamental to linguistics and speech processing. Traditionally, this similarity has been modeled using phonological features, like place and manner of articulation (e.g., is a sound "voiced" or "nasal"?). However, researchers have observed that these feature-based models don't always align with how native speakers actually group sounds. For instance, the way sounds appear together in spoken language (their distributional properties) might play a significant role in shaping our intuitive sense of similarity. This paper investigates whether this distributional information, often overlooked, can explain discrepancies between feature-based models and actual perceptual judgments.
Methods
The core of this research is the creation of a "distributional similarity" model. This model uses a technique inspired by word embeddings (like Word2Vec) but applied to phonemes. Instead of learning word meanings from sentence contexts, the authors learn phoneme representations from their "syllabic environments."
- Data Preparation: A large corpus of English text (the Brown Corpus) was phonemically transcribed. This corpus was then segmented into syllables using a tool that considers English phonotactic rules (rules about how sounds can combine).
- Phoneme Embeddings: A skip-gram model (a type of neural network) was trained on these syllabic environments. The goal was to predict a phoneme based on its surrounding phonemes within a syllable. This process generates multi-dimensional vector representations (embeddings) for each phoneme.
- Similarity Hierarchy: The Euclidean distance between these phoneme embeddings was calculated. A Ward clustering method was then applied to these distances to create a hierarchical representation of phoneme similarity, where phonemes that are closer in the hierarchy are considered more similar.
- Comparison: This distribution-based hierarchy was compared against two other models:
- A feature-based model using phonological features (place, manner, sonority, voicing) as described by Bailey and Hahn [3].
- A perception-based model derived from phoneme confusability data (how often speakers mistake one phoneme for another) from Cutler et al. [10].
Datasets
- Brown Corpus: Used as the source text for generating phonemic data.
- CMU Dictionary: Used for phonemic transcription of words.
- Grapheme-to-phoneme tool: Used for words not found in the CMU dictionary.
- Syllabification tool: Developed by the authors, used to segment phonemic strings into syllables based on English phonotactic constraints.
- Cutler et al. [10] phoneme confusability data: Used to generate the perception-based similarity model.
The paper does not introduce a new benchmark dataset but rather uses existing corpora and experimental data to build and compare models.
Key Insights and Ideas
- Distribution Matters: The study strongly suggests that the frequency and context of phoneme occurrence significantly influence how native speakers perceive their similarity, a factor often underrepresented in traditional feature-based models.
- Perception vs. Features: The perception-based model (derived from confusability) showed closer alignment with the distribution-based model than with the feature-based model in several key areas, highlighting a disconnect between purely linguistic feature descriptions and actual human perception.
- "Natural Classes" Re-evaluated: The distribution-based model offers a data-driven perspective on what constitutes a "natural class" of sounds, suggesting that these groupings might be shaped by usage patterns as much as by inherent phonetic properties.
- Nasal Sounds (M, N, NG): The distribution-based model captures the perceived similarity between M and N, and their dissimilarity from NG, which aligns with half-rhyme usage. This is attributed to NG's unique distributional property of appearing only in syllable codas.
- Palatal Sounds (SH, CH, ZH, JH): The rarity of these sounds in the corpus led to weaker perceived similarity between them in the perception-based model, a finding not directly explained by feature-based models.
- Voicing Contrast: The strong separation between voiced and voiceless phonemes in both the perception-based and distribution-based models underscores the perceptual salience of this feature, which is also reflected in phonotactic constraints.
Key Results
The paper's primary result is the demonstration that a similarity hierarchy generated from phoneme distributional properties (the distribution-based model) exhibits characteristics that align better with a perception-based model (derived from confusability data) than with a traditional phonological feature-based model. Specifically:
- The distribution-based model successfully captures nuanced perceptual similarities, such as the distinction between nasal sounds (M/N vs. NG) and the weaker perceived similarity of rare palatal sounds.
- Both the perception-based and distribution-based models highlight a strong perceptual distinction based on the voicing contrast (e.g., P/B, T/D), which is more pronounced than what is typically captured by feature-based models alone.
Limitations and Future Work
- Limitations:
- The perception-based model is derived from confusability data, which the authors acknowledge may not perfectly represent all aspects of perceptual similarity. Future work aims to gather similarity judgments through more task-specific experiments.
- The current distribution-based model focuses solely on phonemic information and does not incorporate acoustic features.
- Future Work:
- Conduct more task-specific experiments to gather perceptual judgments beyond confusability.
- Combine the distribution-based model with acoustic information to create a purely data-driven model that better captures phonological similarity.
Summary updated