Cultural Emergence of Feature Economy in an Artificial Whistled Language

Tessa Verhoef, Bart de BoerICPhS 2011 · OpenReview

Abstract

Human language is fundamentally combinatorial, relying on a finite set of meaningless building blocks (phonemes) to generate an infinite array of meaningful utterances. Whether this structure is a biological imperative or a product of cultural transmission remains a central question in linguistics. This study employs an experimental iterated learning paradigm using slide whistles to observe the evolution of acoustic signals. Our findings demonstrate that as signals are passed through successive "generations" of learners, they spontaneously transition from holistic, unstructured forms into organized systems that reuse a small set of discrete features—a phenomenon known as feature economy. This transition significantly improves the learnability of the signal set, suggesting that phonological structure emerges as a survival strategy for information transmission.

Introduction

The uniqueness of human speech lies in its phonological organization. Unlike the holistic calls of many non-human animals, human languages are built from a "feature economy," where a limited number of articulatory features (e.g., voicing, place of articulation) are recombined to form various phonemes. While the functional benefits of such a system—efficiency and learnability—are clear, the mechanisms driving its emergence are debated. This research investigates whether cultural transmission alone can carve continuous acoustic space into the discrete, reusable blocks characteristic of human language.

Experimental Methodology

To isolate the effects of cultural evolution from existing linguistic biases, the researchers utilized an iterated learning paradigm. In this setup, the output of one participant serves as the training input for the next, simulating the transmission of language across generations.

The Slide Whistle Apparatus

Participants were tasked with producing sounds using a slide whistle. This instrument was chosen because it allows for a continuous range of pitches but lacks the familiar articulatory constraints of the human vocal tract, thereby preventing participants from simply mapping whistles onto their native phonemic categories.

The slide whistle apparatus used to produce continuous acoustic signals without linguistic bias.

Iterated Learning Procedure

The experiment consisted of four parallel chains, each spanning ten generations. The initial participant in each chain was presented with 12 random, unstructured whistled signals. After a practice phase, they were required to reproduce these signals from memory. Their recorded output then became the training set for the next participant in the chain. Signal similarity was quantified using Derivative Dynamic Time Warping (DDTW) to track changes in pitch and intensity over time.

Results and Analysis

Increased Learnability and Convergence

As the signals progressed through the chains, they became significantly easier to learn. Recall error decreased steadily across generations (Page’s trend test, $p < 0.01$). Simultaneously, the signals within each set became more similar to one another, converging on a shared set of acoustic building blocks. The average nearest neighbor distance between whistles in a set decreased significantly ($p < 0.001$), indicating a shift from a diverse, holistic set to a more uniform, structured system.

Emergence of Feature Economy

The most striking result was the spontaneous emergence of "feature economy." Participants began to reuse specific acoustic elements—such as a particular pitch slide or a steady-state note—across multiple signals in their repertoire.

Evolution of whistled signals across generations, highlighting the reuse and regularization of specific pitch features.

This regularization was not merely a loss of information but a productive strategy. By the 10th generation, the random whistles had transformed into predictable systems where elements were reused in "language-specific" ways. For instance, one chain might evolve to use only descending slides, while another might favor specific rhythmic patterns. This structure allowed participants to group whistles into subsets, making the entire set of 12 signals much easier to memorize and reproduce.

Emergent Co-articulation

The study also observed the emergence of co-articulation, where the terminal pitch of one segment influenced the onset of the next. This resulted in more fluid, "speech-like" transitions, further suggesting that the pressures of production and perception drive signals toward structured forms.

Discussion

The results provide strong evidence that the combinatorial structure of language can emerge through the process of cultural transmission. The drive for learnability acts as a selective pressure, favoring signal sets that are organized around a small number of reusable features.

While this study focused on a small vocabulary of 12 signals without assigned meanings, the observed trends suggest that larger vocabularies would likely face even greater pressure to become structured. Future research should explore how the need to communicate specific meanings interacts with these pressures for learnability. Furthermore, the fact that similar structural emergence has been observed in gestural modalities suggests that the drive for feature economy is a general cognitive principle of human learning rather than a modality-specific acoustic phenomenon.

Conclusion

By simulating the "evolution" of whistled sounds, this research demonstrates that the building blocks of speech are not necessarily hard-wired into our biology. Instead, they can emerge naturally as a consequence of the human brain's attempt to learn, remember, and pass on complex information. Feature economy is thus a fundamental property of culturally transmitted systems, providing a bridge between continuous signals and discrete linguistic structure.

Summary updated

BibTeX
@inproceedings{verhoef2011cultural,
  author = {Tessa Verhoef and Bart de Boer},
  title = {{Cultural Emergence of Feature Economy in an Artificial Whistled Language}},
  booktitle = {ICPhS},
  year = {2011},
  url = {https://openreview.net/forum?id=Gce30oSfGc},
}
Figure