| 研究実績の概要 |
In cocktail party scenarios, many information need to be explored in order to identify different speech (or sound) sources, in particular, who is speaking (speaker information) is one of the most important information for identifying speech sources. In order to combine advantages of both discriminative and generative classifier models for speakers, we proposed to couple a generative model in a discriminative learning for speaker recognition. Our framework showed a large improvement compared with state of the art models.
|
