Recovering task dynamics from speech acoustics: numerical results and the application of the method in speech technology.
| Number | 919 |
|---|---|
| Year | 1994 |
| Drawer | 17 |
| Entry Date | 07/16/1998 |
| Authors | McGowan, R.S. |
| Contact | |
| Publication | Proceedings of Technical Conference on Telecommunications R & D in Massachusetts. University of Massachusetts, Lowell, MA. pp. 33-44. |
| url | http://www.haskins.yale.edu/Reprints/HL0919.pdf |
| Abstract | The problem of recovering speech articulation from the speech acoustic signal, the speech inverse problem, has seen much progress in the last forty years. Much of this progress has resulted from increased knowledge of speech motor control, which may allow the mapping from acoustics to articulation to be constrained sufficiently to be unique and robust. The task-dynamic model of speech production, developed at Haskins Laboratories by Saltzman and his colleagues, incorporates knowledge of speech motor control into a computational model. Task-dynamics is used in the method proposed here as a means to constrain the inverse mapping. A brief description of the proposed method is given, including task-dynamics and the genetic algorithm that is used in the optimization. Then some results from computer experiments are given. The method is analysis-by-synthesis, where the speech of a proposed articulation is compared to the speech data. The proposed articulations are specified by task-dynamic parameters as coded into chromosomes strings. A genetic algorithm is applied to a population of these strings, so that the speech of the proposed articulations approaches that of the data. Articulatory recovery, or at least articulatory constraint, has been proposed as a way to perform bit-rate reduction and as part of the way to do automatic speech recognition. future uses of this method in automatic speech recognition are proposed. |
| Notes |
Search Publications