VFS speaker adaptation with transfer vectors selected by reliability according to the amount of training data
✍ Scribed by Kazumi Ohkura; Masayuki Iida
- Publisher
- John Wiley and Sons
- Year
- 1998
- Tongue
- English
- Weight
- 495 KB
- Volume
- 29
- Category
- Article
- ISSN
- 0882-1666
No coin nor oath required. For personal study only.
✦ Synopsis
Proposed in this paper is a method for speaker adaptation, namely, transfer vector field smoothing (VFS) with respect to transfer vector reliability. We also discuss the results obtained in applying the proposed method to MAP-VFS speaker adaptation. The transfer vectors are maximum a posteriori probability (MAP) estimates obtained using maximum likelihood estimation with weights according to the data distribution. These weights may be considered as transfer vector probabilities. With VFS, transfer vectors for interpolation and smoothing are found by using the k nearest neighbors rule from among all transfer vectors related to trained distributions. With the proposed method, however, the k nearest neighbors rule is applied to high-reliability transfer vectors selected according to the amount of training data in the course of MAP estimation. This ensures more active use of vector reliability when implementing transfer vector field smoothing. Word recognition experiments were performed to show that the proposed method supports the same recognition performance as does the conventional method (without transfer vector selection) while using only half the training data.