 |
|
|
|
|
|
DIVA
DIVA (Directions Into Velocities of Articulators) is a neural network model of speech
motor skill acquisition and speech production. In computer simulations, the model learns to control
the movements of a computer-simulated vocal tract in order
to produce sequences of phoneme strings. The model's neural mappings are tuned during a babbling
phase in which auditory feedback from self-generated speech sounds is used to learn the relationship
between motor actions and their acoustic consequences. After learning, the model can produce
arbitrary combinations of phonemes, even in the presence of constraints on the articulators.
DIVA provides unified explanations for a number of
long-studied speech production phenomena including motor equivalence, contextual variability,
speaking rate effects, anticipatory coarticulation, and carryover coarticulation.
The model is schematized in the figure below. Each block in the diagram corresponds to a hypothesized
set of neurons in the human speech system.
Figure 1: Schmatic of the DIVA Model
The first set of synaptic weights (the filled semi-circle between the speech sound map and
the planning direction vector) encodes auditory and orosensory targets for phonemes learned
during babbling. The learned speech sound targets take the form of multi-dimensional regions, rather
than points, in auditory and orosensory spaces. The second set of weights
(labeled directional mapping transform the desired movement
direction in auditory space into movement directions in articulator space.
The problem of mapping from auditory or orosensory space to articulator
space is ill-posed, and the model's mapping is related to the
Moore-Penrose pseudoinverse of the Jacobian matrix relating the auditory,
somatosensory and articulatory spaces. The final mapping, labeled
forward model transforms orosensory feedback from the vocal tract
and an efference copy of the motor commands into an auditory
representation that can be used to control speech movements without
relying on auditory feedback.
DIVA Papers:
DIVA Demos
Hypothesized Neural Correlates of
DIVA
|
|
|
|
