Learning the Mapping and Targets HBR 250798

Shown below is an animation sequence of hidopt learning both the MLP mapping and the targets for different vowels in the vowel sequence shown. It can be seen that in the early stages of training the MLP first approximates the mapping from hidden space to the surface acoustic space by a linear mapping, as the spectra for diffent segments simply fade in and fade out. In the later stages of training, the transitions between segments are learnt, and continuity can be seen in the formants between segments.