Speaker Recognition System with Limited Data based on LightGBM and Fusion Features
Yuan Xiwen, Xiaosong Zhu · 2021
The field of speaker recognition has recently developed rapidly, while the accuracy of speaker recognition has generally been declining due to the limited data. Considering the inadequacy of the single feature parameter, the paper selects five common features and gets a new feature M-IM-G-L-BFCC by fusing them. LightGBM, an advanced ensemble learning algorithm in machine learning, is then selected for pattern recognition and a large number of experiments are carried out on different datasets. The last simulation results show that the new fused feature parameters can represent more features, which has greatly improved the recognition accuracy of LightGBM.