Investigating the Emergence of Speech Sounds

Bart G. de BoerIJCAI 1999

Introduction

Human languages exhibit remarkably similar vowel structures, often gravitating toward symmetric systems like the common [a, e, i, o, u] configuration. For decades, the origin of these patterns has been a subject of intense debate. While some linguists propose an innate "language organ," others suggest that these structures are products of cultural evolution. This paper investigates whether realistic speech sounds can emerge spontaneously through local interactions among agents without a central designer or explicit optimization.

Methods: The Imitation Game

The study utilizes a multi-agent simulation where agents interact through "Imitation Games." Each agent possesses an articulatory synthesizer and an associative memory for vowel prototypes. During a game, an Initiator produces a sound based on its internal prototypes, adding acoustic noise to simulate real-world variability. An Imitator then attempts to match this sound using its own repertoire. Success is determined by whether the Initiator recognizes the Imitator's response as the original vowel.

Agents update their repertoires based on these interactions: shifting prototypes toward successful imitations, pruning unsuccessful ones, and occasionally introducing random new vowels. This process models a dynamical system where global structure emerges from local pressures for communicative success.

Evolution of vowel systems over 25, 500, 1000, and 5000 imitation games.

Key Insights and Results

The simulation demonstrates that coherent vowel systems act as "attractors" in the acoustic space. Even without a global objective, the population self-organizes into stable, symmetric configurations.

Symmetry and Natural Inventories

The emergent systems closely mirror real-world data from the UPSID (UCLA Phonological Segment Inventory Database). In simulations resulting in five-vowel systems, 88% were symmetric, nearly matching the 89% observed in human languages. The agents spontaneously developed vowels corresponding to [i], [e], [a], [o], and [u], which are the most frequent vowels globally.

A representative emergent vowel system showing distinct clusters in the F1/F2' acoustic space.

The Role of Noise and Cultural Transmission

Contrary to intuition, acoustic noise is beneficial; it prevents vowel clusters from collapsing into points, maintaining realistic spacing. Furthermore, the systems proved robust to population turnover. As older agents were replaced by new ones, the vowel structures persisted, demonstrating that speech sounds can be preserved through cultural transmission alone.

Discussion

While the simulation successfully predicts the emergence of symmetry and common inventories, it slightly underestimates the prevalence of five-vowel systems, often peaking at four. This discrepancy likely stems from limitations in the current perception model. Nonetheless, the research provides strong evidence that the fundamental building blocks of speech can arise from self-organization and cultural evolution, rather than biological hard-wiring. Future work will explore applying these principles to more complex linguistic structures like consonants and grammar.

Summary updated

BibTeX
@inproceedings{boer1999investigating,
  author = {Bart G. de Boer},
  title = {{Investigating the Emergence of Speech Sounds}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1999},
}
Figure