Abstract
This paper explores the hypothesis that the structured stages of early vocal development in infants can emerge from a general mechanism of intrinsically motivated spontaneous exploration, or "curiosity-driven learning." We present a computational model where a vocal agent autonomously structures its own learning schedule by seeking to maximize "competence progress." The model utilizes a realistic physical simulation of the vocal tract and auditory system. Results demonstrate that the agent self-organizes into developmental stages—moving from simple phonation to unarticulated sounds and finally to articulated babbling—that closely mirror infant psychology. Furthermore, the model explains the transition from solitary vocal play to social imitation as an emergent property of the drive for learning progress.
Introduction
Human infants follow a remarkably consistent path in vocal development. They start with "non-speech" sounds such as squeals and growls, move to vowel-like quasi-vowels, then to canonical babbling (proto-syllables), and finally to words influenced by their environment.
Historically, computational models of speech acquisition have "hard-coded" these transitions, assuming a biological clock or a pre-set curriculum. This paper challenges that assumption, investigating whether "curiosity-driven learning"—a drive to explore activities where one is improving the fastest—can explain both the sequence of vocal skills and the eventual shift toward social imitation.
Methods
The researchers developed a "curious" vocal agent comprising several integrated components:
- Vocal Tract & Auditory Model: A version of the DIVA/Maeda model, providing a realistic physical simulation of the human vocal tract.
- Motor Control: The agent controls 7 articulator parameters and 2 phonation parameters. An over-damped spring-mass system simulates the physics of movement, allowing for "co-articulation," where the position of the jaw or tongue in one command affects the starting point of the next.
- Intrinsic Motivation (SAGG-RIAC & SGIM-ACTS): This "brain" of the model performs goal-directed exploration. Instead of random babbling, the agent picks a desired auditory goal and tracks its "competence progress." It focuses its attention on regions of the auditory space where its ability to reach goals is improving most rapidly—its "progress niches."
Key Insights
- Stages as "Attractors": The developmental sequence (No Phonation → Unarticulated → Articulated) emerged in almost all simulations. This was not programmed; it occurred because simple sounds provide "easy" early progress. Once phonation was mastered, the "progress niche" shifted to more complex articulated sounds.
- The Logic of Social Shift: The agent initially ignores social peers (the "Teacher"). Because the teacher's sounds are too complex, the agent makes zero progress trying to imitate them. It only begins to imitate once it has "leveled up" its motor skills through self-exploration, making the teacher's sounds a new, viable progress niche.
- Self-Organization of Complexity: The agent naturally moves from 0D auditory subspaces (silence) to 3D (unarticulated vowels) to 6D (articulated syllables) as it exhausts the learning potential of simpler sounds.
Results
The model's behavior closely mirrored the developmental timelines described in infant psychology. Across 9 independent simulations, the agent consistently spent the first ~30,000 vocalizations learning to control phonation before spending the next ~120,000 vocalizations on unarticulated sounds and finally moving to articulated babbling. In social trials, the agent transitioned to imitation around the 450th model update—a shift that was purely emergent, driven by the internal drive to find the highest rate of improvement.
Discussion and Future Work
While the model successfully replicates developmental stages, it faces limitations. It assumes the agent can map a teacher's auditory formants to its own (the correspondence problem) and uses a simplified auditory representation. Future work should explore higher-dimensional motor models and the interaction between physical maturation (a growing vocal tract) and curious learning. This research suggests that the complex, staged nature of vocal development does not require a specific biological clock for each stage, but can instead be explained by a single, general-purpose drive for curiosity.
Summary updated