Zvýšení spolehlivosti v automatickém rozpoznávání řečníka
Project Period: 1. 1. 2017 - 31. 12. 2019
Project Type: grant
Agency: Czech Science Foundation
Program: Juniorské granty
automatic speaker recognition;robustness;adaptation;speech
Speaker recognition systems have gained very high recognition performance in the recent years. However, it has been shown that system performance degrades when the recognition data domain differs from the one used for system parameter training. Also, introducing additive noise (e.g. background traffic noise), convolutive noise (e.g. reverb of the room), or channel noise (e.g. telephone codec) to the recording further degrades the performance. The solutions to these issues are to a) seek for techniques for robust modeling, and b) to develop methods for system adaptation. In this project, we want to focus on both of these approaches.