A Deep Generative Model of Vowel Formant Typology

Ryan Cotterell, Jason Eisner

Abstract

Ryan Cotterell, Jason Eisner. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers).

2018.

Introduction

Linguistic typology seeks to understand why certain language structures are more prevalent across human cultures than others. In the domain of phonology, a striking observation is that vowel systems are not random; almost every language contains the vowel /i/, while very few possess /y/. Traditionally, these patterns were studied using discrete symbols like the International Phonetic Alphabet (IPA). However, IPA symbols are abstractions that can obscure fine-grained acoustic variation.

This research introduces a deep generative model to explain vowel system typology by modeling the actual resonant frequencies of the human vocal tract—known as formants ($F_1$ and $F_2$). The goal is to capture two fundamental linguistic principles: dispersion (vowels should be distinct) and focalization (vowels should target specific acoustic "sweet spots").

The distribution of vowel formants across languages in the $F_1$ and $F_2$ acoustic space.

The Generative Model

The authors propose a Bayesian framework that operates on raw acoustic data. The model consists of four primary components:

  1. Universal Phone Discovery: A latent set of "universal phones" is represented as Gaussians in a 2D space. The number of phones is sampled from a Poisson distribution.
  2. Modeling Dispersion with DPPs: To prevent vowels within a single language from overlapping, the model employs a Determinantal Point Process (DPP). This probabilistic framework favors diverse subsets of items, effectively learning the optimal spacing between vowels from the data.
  3. Modeling Focalization: A neural network assigns a "focalization score" to regions of the acoustic space, representing the inherent desirability of certain sounds regardless of their neighbors.
  4. Neural Acoustic Warping: A diffeomorphism—a reversible, differentiable neural transformation—warps the latent "ideal" phones into the observed $F_1, F_2$ space. This captures the non-linear relationship between mental targets and acoustic outputs.

Experimental Setup

The study uses the Becker-Kristal (2010) corpus, comprising phonetic data from 233 distinct languages. Each entry includes a language's vowel inventory and their corresponding first two formant values. The performance of the unsupervised deep generative model is compared against a "Supervised" baseline that relies on standard IPA labels.

Key Results

The model was evaluated using cross-entropy (measuring model surprise) and a "cloze task" (predicting missing vowels in a partial inventory).

Comparison of model performance across different numbers of universal phones ($N$) and architectural variants.

The results demonstrate several key findings:

  • DPP Superiority: Models incorporating dispersion (DPP) consistently outperformed those treating vowels as independent (Bernoulli Point Process).
  • Deep Learning Impact: The inclusion of the neural transformation ($\nu_\theta$) significantly improved performance over linear models, indicating that simple Gaussian distributions cannot capture the complex geometry of the acoustic space.
  • IPA Limitations: The supervised model performed poorly compared to the deep generative model, suggesting that IPA labels are often "leaky abstractions" that fail to capture true underlying acoustic typology.

Discussion and Future Work

The success of this model highlights that dispersion is a statistical prior learned from data rather than a hard-coded rule. Furthermore, the necessity of neural warping suggests that the mapping between mental vowel targets and acoustic outputs is complex, likely due to the physics of the vocal tract.

Future research could extend this work by incorporating higher-order formants ($F_3$), vowel length, and tone. Additionally, accounting for language families (phylogeny) through hierarchical modeling could provide a more nuanced understanding of how vowel systems evolve.

Summary updated

BibTeX
@inproceedings{cotterell2018deep,
  author = {Ryan Cotterell and Jason Eisner},
  title = {{A Deep Generative Model of Vowel Formant Typology}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
  url = {https://openreview.net/forum?id=zWc7UqQ-r8x},
}
Figure