| 摘要 |
PROBLEM TO BE SOLVED: To provide a speaker model creation apparatus that creates a standard speaker model for increasing recognition accuracy. SOLUTION: The speaker model creation apparatus comprises a clean learning data storage part 131 for storing clean features representing speech features in a noiseless environment, a noisy learning data storage part 132 for storing noisy features representing speech features in a noisy environment, a transformation parameter calculation part 110 for calculating first model parameters representing parameters of a speaker model maximizing a first likelihood of the clean features, and calculating transformation parameters for transforming the distribution of each speaker's clean features into a distribution represented by a speaker model of the first model parameters which maximizes the first likelihood, and a model parameter calculation part 120 for transforming each speaker's noisy features with the transformation parameters, and calculating second model parameters representing parameters of a speaker model maximizing a second likelihood of the transformed noisy features. COPYRIGHT: (C)2009,JPO&INPIT
|
