Data-driven automated acoustic analysis of human infant vocalizations using neural network tools
Anne S Warlaumont1, D Kimbrough Oller, Eugene H Buder
1School of Audiology and Speech-Language Pathology, The University of Memphis, 807 Jefferson Avenue, Memphis, Tennessee 38105, USA. awarlmnt@memphis.edu
The Journal of the Acoustical Society of America
|April 8, 2010
Summary
This study introduces a novel neural network approach for analyzing infant vocalizations, moving beyond traditional acoustic measures. The method successfully classifies vocalizations by type, age, and speaker, offering a new tool for developmental research.
Area of Science:
- Speech and Language Pathology
- Computational Linguistics
- Developmental Psychology
Background:
- Traditional acoustic analysis of infant vocalizations relies on measures from adult speech.
- These methods may not fully capture the unique characteristics of early vocal development.
Purpose of the Study:
- To propose and evaluate an alternative method using data-derived spectrographic features for infant vocalization analysis.
- To employ a neural network for classifying prelinguistic infant utterances and identifying speaker and age information.
Main Methods:
- Analysis of 1-second spectrograms from six infants (3-11 months) using a self-organizing map and a single-layer perceptron neural network.
- The self-organizing map generated holistic, data-derived spectrographic receptive fields.
- The single-layer perceptron was trained for classification tasks: phonatory category, infant age, and infant identity.
Main Results:
- Classification performance for all three tasks (category, age, identity) was significantly better than chance.
- The neural network approach demonstrated effectiveness in extracting meaningful features directly from vocalization data.
- Comparison with a fully supervised multilayer perceptron indicated the viability of the proposed architecture.
Conclusions:
- The proposed neural network method offers a complementary approach to traditional acoustic analysis for infant vocalizations.
- This data-driven technique facilitates the derivation of holistic spectrographic features and automatic classification of infant speech.
- The findings support the utility of such tools in understanding early vocal development and prelinguistic communication.
