Abstract:

The articulatory codebook approach to voice mimic systems is complicated by large amounts of data and expensive acoustic comparisons when searching the codebook. If, during codebook construction, the acoustic parameters are quantized and mapped back to the articulatory space, the inverse is accomplished without acoustic comparisons. Since the inverse is nonunique, articulatory trajectories must also be estimated to resolve acoustic parameters which map to multiple model shapes. This can be managed with a tracking technique that uses dynamic properties (position and velocity) of each articulatory parameter to estimate the next model shape. A recurrent algorithm results which finds an optimal path through the model shape variations. These approaches are helpful, but they do not address the problem that codebooks may not cover the entire articulatory space. Furthermore, there is no performance measure which indicates the extent of the coverage. This limitation is overcome by an analytic inverse mapping, rather than a table look-up, which relates the first two formant frequencies with articulatory parameters. This relationship, which is well established for vowels, uses a distinctive region model (DRM) to approximate three area function parameters as a function of the first two formant frequencies. [Research supported by ARPA DAST 63-93-C-0064.]