Data-driven automated acoustic analysis of human infant vocalizations using neural network tools

摘要

Acoustic analysis of infant vocalizations has typically employed traditional acoustic measures drawn from adult speech acoustics, such as f0, duration, formant frequencies, amplitude, and pitch perturbation. Here an alternative and complementary method is proposed in which data-derived spectrographic features are central. 1-s-long spectrograms of vocalizations produced by six infants recorded longitudinally between ages 3 and 11 months are analyzed using a neural network consisting of a self-organizing map and a single-layer perceptron. The self-organizing map acquires a set of holistic, data-derived spectrographic receptive fields. The single-layer perceptron receives self-organizing map activations as input and is trained to classify utterances into prelinguistic phonatory categories (squeal, vocant, or growl), identify the ages at which they were produced, and identify the individuals who produced them. Classification performance was significantly better than chance for all three classification tasks. Performance is compared to another popular architecture, the fully supervised multilayer perceptron. In addition, the network’s weights and patterns of activation are explored from several angles, for example, through traditional acoustic measurements of the network’s receptive fields. Results support the use of this and related tools for deriving holistic acoustic features directly from infant vocalization data and for the automatic classification of infant vocalizations.