A Neural Model of Speech Production

This paper describes the most recent version of the DIVA model, a neural network model of the brain computations underlying the acquisition and production of speech sounds. The model, which is implemented as a set of equations representing neural activity and synaptic strengths, is designed to account for the results of functional magnetic resonance imaging (fMRI) and electromagnetic midsagittal articulometry (EMMA) experiments concerning the production of speech. Computer simulations of the model have been performed to illustrate its ability to account for speaker-specific articulator movements in different phonetic contexts, as well as fMRI activations seen during normal and perturbed speech. The model is also used to generate predictions that guide new fMRI and EMMA experiments aimed at achieving a better understanding of the neural bases of speech. The results of these experiments are in turn used to further refine the model. Finally, the model can be used to investigate the effects of various types of neurological damage on speaking skills. INTRODUCTION: OVERVIEW OF THE DIVA MODEL Figure 1 provides an overview of the DIVA model (e.g., Guenther, 1994; Guenther et al., 1995; 1998; Perkell et al., 2000; Callan et al., 2000). The model consists of a neural network controller whose cells correspond to boxes and synaptic weights correspond to arrows in the figure. The neural network utilizes a babbling stage to learn the neural mappings (arrows) necessary for controlling an articulatory synthesizer (e.g., Maeda, 1990). The output of the model (labelled “To Muscles” in Figure 1) specifies the positions of the 7 articulators that determine the vocal tract shape in the articulatory synthesizer. Speech Sound Map (Premotor Cortex) Articulator Velocity and Position Maps (Motor Cortex) Auditory Error Map (Auditory Cortex) Somatosensory Error Map (Somatosensory Cortex) Auditory Goal Region Somatosensory Goal Region Somatosensory State Auditory State