| US 7,519,531 B2 | ||
| Speaker adaptive learning of resonance targets in a hidden trajectory model of speech coarticulation | ||
| Alejandro Acero, Bellevue, Wash. (US); Dong Yu, Kirkland, Wash. (US); and Li Deng, Sammamish, Wash. (US) | ||
| Assigned to Microsoft Corporation, Redmond, Wash. (US) | ||
| Filed on Mar. 30, 2005, as Appl. No. 11/93,833. | ||
| Prior Publication US 2006/0229875 A1, Oct. 12, 2006 | ||
| Int. Cl. G10L 19/06 (2006.01) | ||
| U.S. Cl. 704—209 [704/255] | 18 Claims |
| 1. A computer-implemented method of training a hidden trajectory model, of a speech recognition system, which generates Vocal
Tract Resonance (VTR) targets, the method comprising:
obtaining initial VTR target parameters using speaker-independent training;
storing the initial VTR target parameters as generic VTR target parameters;
iteratively updating the generic VTR target parameters, for each of a plurality of speakers in a training set, using speaker-adaptive
training to generate updated generic VTR target parameters, wherein iteratively updating the generic VTR target parameters,
for each particular speaker of the plurality of speakers in the training set, further comprises iteratively scaling the generic
VTR target parameters using a corresponding speaker-dependent scaling factor for the particular speaker to generate the updated
generic VTR target parameters;
storing the updated generic VTR target parameters on a computer storage medium, for use by a target selector to generate VTR
sequences, as the generic VTR target parameters in order to configure the hidden trajectory model to perform speech recognition;
and
performing speech recognition using the configured hidden trajectory model.
|