A Method of Speaker Recognition for Small-scale Speakers Based on One-versus-rest and Neural Network
Qingyun Sun, Linkai Luo, Hong Peng, Chaojie An · 2019
There are a large number of application for the speaker recognition with a small size of speakers. A characteristic for this type of application is that the number of speakers is often changed slightly. To adapt this characteristic, we apply one-versus-rest(OvR) and neural network to implement speaker recognition. The task is first split into a series of binary classification problems by OvR strategy. The binary classification problems are then solved by neural network and the recognition result is obtained by the simple majority voting. When a few speakers are added to the set of speakers, we can only train the binary classifiers for the newly added speakers. The experiment on a dataset with small-scale speakers shows that both the recognition accuracy and the time cost is satisfactory, which indicates that the method is effective and feasible.