A pipeline for stochastic and controlled generation of realistic language input for simulating infant language acquisition.

Abstract

Computational models of early language development involve implementing theories of learning as functional learning algorithms, exposing these models to realistic language input, and comparing learning outcomes to those in infants. While recent research has made major strides in developing more powerful learning models and evaluation protocols grounded in infant data, models are still predominantly trained with non-naturalistic input data, such as crowd-sourced read speech or text transcripts. This is due to the lack of suitable child-directed speech (CDS) corpora in terms of scale and quality. In parallel, the question of how properties and individual variability in language input affect learning outcomes is an active area of empirical research, underlining the need for realistic yet controllable data for modeling such phenomena. This paper presents a solution to the training data problem through stochastic generation of naturalistic CDS data using statistical models, thereby enabling controlled computational simulations with naturalistic input. We provide a proof-of-concept demonstration of the approach by showing how naturalistic CDS transcripts can be generated with a language model conditioned on recipient information (here, infant age), and how text-to-speech systems can be used to convert the transcripts to high-quality speech with a controllable speaking style. We also conduct modeling experiments with generated speech corpora by varying different aspects of the data, showing how this maps into different learning outcomes, thereby demonstrating the feasibility of the approach for controlled language learning simulations. Finally, we discuss the limitations of using synthetic data in general, and of the present proof-of-concept pipeline in particular.