|
During the first year of life, human infants practice controlling the muscles of the vocal tract. They also hear their own vocalizations as well as those of others, such as their parents and siblings. This motor practice and perceptual experience leads to changes in the connections between neurons throughout the infant's nervous system, laying the foundation for later speech and language.
I have developed an artificial neural network model that listens to audio examples of real infant vocalizations. With experience, the connections in the network are gradually adjusted, changing the receptive fields of the neurons. These learned receptive fields become organized topographically, so that neighboring nodes in the network prefer similar sounding inputs. This is similar to the organization in real nervous systems. The model's responses to infant vocalization inputs are validated by comparison to humans' judgments regarding the same infant vocalizations. I am now working to extend this work in order develop a model of the infant that both perceives the sounds produced by itself and by its caregivers and produces sounds by activation motor neurons that in turn drive a speech synthesizer that simulates the human vocal tract.
Even realistic neural network models usually represent extreme simplifications of the real physical system. In addition, the scope of behaviors (in this case, vocalizations) that the models produce or perceive is relatively small. One reason for this is that these modeling projects are fairly young and the system being modeled is very complex. Another reason is that neural network models and realistic speech synthesizers tend to be computationally expensive. Technical advances and high-performance computing will permit improvements such as increases in network size, improvement in temporal dynamics, and increases in the amounts and realism of vocalization data perceived and produced.
Warlaumont, A. S., Oller, D. K., Buder, E. H., Kozma, R., & Dale, R. Data-driven automated acoustic analysis of human infant vocalizations using neural network tools. Journal of the Acoustical Society of America, vol. 127, pp. 2563-2577, 2010.
A. S. Warlaumont and R. Dale, "The missing chapter: The interaction between behavioral and symbolic inheritance," Behavioral and Brain Sciences, vol. 30, pp. 377-378, 2007.
A. S. Warlaumont, "Effects of contextual regularities on memory for a configuration of objects within a scene," Psychology honors thesis, Cornell University, 2005.
|
