In this paper we have shown that inverse modeling techniques can be used to map representations of sound to physical parameters for sound-generating models. The inverse modeling strategy depends on the geometry of the solution space. If the solution region is convex we can use a direct inverse-modeling strategy, such as back-propagation in a two-layer, feed-forward network. However, non-convex solution regions require a more sophisticated approach to deriving the inverse model. One such approach is that of using distal teachers with forward models. We implemented such a system and obtained satisfactory results for recovering physical parameters for models of violin strings.\
With careful implementation, the forward modeling strategy is general enough to be applied to many inverse modeling problems in the auditory domain. We are currently expanding the scope of the current research to include models of other sounding systems; e.g. single-reed, brass and vocal-tract models. The outputs of these inverse models can be treated as features to which we can apply pattern recognition techniques for source classification and gesture recognition. An example of this is to use parameters recovered from real musical performances to classify different playing styles or performance techniques, perhaps creating a machine listening system that can understand the subtleties of musical performance.\
Future work will include the development of distance functions for
auditory data that take into account human perceptual factors.
Time-domain representations are unsatisfactory for many applications
of audio inverse modeling; there are many different time-domain
representation of a signal that produce a single auditory percept.
This is due, in large part, to the low salience of phase in the human
auditory system.
We have experimented with a constant-Q frequency representation which
better represents the perceptual distance between auditory stimuli.
\