A Deep Generative Model of Vowel Formant Typology

Ryan Cotterell, Jason Eisner

Abstract

What makes some types of languages more probable than others? For instance, we know that almost all spoken languages contain the vowel phoneme /i/; why should that be? The field of linguistic typology seeks to answer these questions and, thereby, divine the mechanisms that underlie human language. In our work, we tackle the problem of vowel system typology, i.e., we propose a generative probability model of which vowels a language contains. In contrast to previous work, we work directly with the acoustic information -- the first two formant values -- rather than modeling discrete sets of phonemic symbols (IPA). We develop a novel generative probability model and report results based on a corpus of 233 languages.

Introduction

Human languages exhibit striking universal patterns in their phonological structures. In linguistic typology, it is a well-known universal that all spoken languages possess vowels, and their selection is typically governed by two core principles: focalization (a preference for acoustically stable, "quantal" vowels like /i/ or /u/) and dispersion (the tendency for vowels to be acoustically distinct to minimize perceptual confusion for the listener).

Traditional computational models of vowel typology have relied on discrete International Phonetic Alphabet (IPA) symbols. However, IPA symbols are abstractions that mask significant cross-linguistic acoustic variation—for example, the /i/ in English differs subtly from the /i/ in German. This research addresses this gap by modeling vowel inventories directly in a continuous 2D acoustic space defined by the first two formants ($F_1$ and $F_2$), which are the resonant frequencies of the vocal tract determining vowel quality.

Methodology

The proposed Bayesian generative model simulates how languages select their vowel inventories through a multi-stage process:

  1. Universal Phones: The model first samples a universal set of $N$ Gaussian "phones" in a latent 2D space.
  2. Language Inventories via DPP: Subsets of these phones are sampled for individual languages using a Determinantal Point Process (DPP). The DPP is the core innovation, mathematically encoding both dispersion (repulsion between similar phones) and focalization (baseline preferences for certain high-quality phones).
  3. Acoustic Realization: Latent phones are mapped to observed acoustic formants ($F_1, F_2$) via a neural network-based diffeomorphism. This invertible mapping accounts for the non-linear warping between the human perceptual/articulatory space and raw acoustic frequencies.

Training is performed using MAP-EM (Maximum A Posteriori Expectation-Maximization), utilizing an MCMC Gibbs sampler for the E-step and reversible-jump MCMC to determine the optimal number of universal phones ($N$).

Data and Representation

The model was evaluated using the Becker-Kristal (2006) Corpus, which contains vowel inventory data for 233 distinct languages. Each vowel in the dataset is represented by its $F_1$ and $F_2$ values measured in Hertz.

Vowel formant distribution in the F1/F2 acoustic space across multiple languages, showing the continuous nature of phonetic data.

By operating in this continuous space, the model captures phonetic nuances that discrete IPA labels often lose. The unsupervised clustering of this data suggests that field linguists may often force-fit acoustic realities into standard IPA categories, losing valuable phonetic detail in the process.

Key Results

The full model (DPP + Neural Network) significantly outperforms baseline approaches across several metrics, including cross-entropy and cloze task performance.

Quantitative comparison of model performance across different configurations (DPP vs BPP) and phone counts (N).

Key findings include:

  • Component Impact: Removing the DPP (ignoring dispersion) or the neural network (ignoring non-linear warping) led to substantial performance degradation. For $N=50$, the full model achieved 222.85 nats, compared to 231.70 without DPP and 320.05 without the neural network.
  • Supervised Baseline Failure: A model forced to use supervised IPA labels performed significantly worse (1610.37 nats), highlighting the inconsistency of discrete phonetic transcription across different languages.
  • Predictive Accuracy: In cloze tasks—predicting missing vowels in a language's inventory—the full model consistently achieved the lowest expected Euclidean distance error in the formant space.

Discussion and Future Work

This work demonstrates that a continuous, generative approach to vowel typology provides a more accurate and theoretically grounded framework than discrete symbol-based models. However, limitations remain. The current model assumes languages are drawn independently, ignoring historical and phylogenetic relationships. Additionally, it focuses solely on $F_1$ and $F_2$, omitting features like lip rounding ($F_3$), tone ($F_0$), and duration. Future research could integrate these features and explore the evolutionary mechanisms driving the selection of optimal vowel systems.

Summary updated

Abstract

What makes some types of languages more probable than others? For instance, we know that almost all spoken languages contain the vowel phoneme /i/; why should that be? The field of linguistic typology seeks to answer these questions and, thereby, divine the mechanisms that underlie human language. In our work, we tackle the problem of vowel system typology, i.e., we propose a generative probability model of which vowels a language contains. In contrast to previous work, we work directly with the acoustic information -- the first two formant values -- rather than modeling discrete sets of phonemic symbols (IPA). We develop a novel generative probability model and report results based on a corpus of 233 languages.
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