A training set comprising parameters to two different physical models, brass and reeds, was constructed by sparsely sampling the physical parameter space in a similar way to experiment one. However, we added two class tags at the start of the parameter vector indicating which of two physical models should be chosen for the given set of parameters. A one indicated selection and a zero indicated rejection. The values of these units at the output of the estimator can be interpreted as likelihood estimates. Thus we implemented a simultaneous classification and estimation system. We used the same brass model as described in Section 5.1 and added a single-reed physical model. The reed model was configured to synthesize the sound of an alto saxaphone. The parameters to the single reed model are shown in Table 3.\
Table 3: Physical Model Parameters for Single-Reed Model
As in the first experiment, a forward model was estimated followed by a direct inverse model. Finally the distal inverse model was trained to a criterion of 0.001 as in the above experiment.