一种鲁棒性的话者确认阈值的设计方法

A METHOD FOR DESIGNING ROBUST VERIFICATION THRESHOLD

本文提出了一种基于子带技术和人工神经网络技术的鲁棒性的话者确认阈值的设计方法,将语音信号的有效频段划分为几个子带独立地训练或识别,并在对各个子带的输出数据融合的基础上作最后的判决。各个子带的模型训练及识别采用矢量量化技术,数据的融合技术则采用BP型人工神经网络。采用子带技术可以提高话者确认阈值的时间鲁棒性,采用神经网络技术一方面是为了对各子带的输出进行非线性数据融合,另一方面则是为了能够对话者本人的数据和冒认者的数据进行混合训练,以使训练出的确认阈值对冒认者的不确定性具有鲁棒性。本文提出的设计方法可得到鲁棒性的确认阈值,并得到了实验验证。

A method for the designing of robust verification threshold based on sub - band and ANN techniques is presented in this paper. The whole valid band is divided into several sub - bands to train and classify independently. The final decision is made based on the merging of the resultes from these sub - bands. VQ technique is used for the modeling of the sub - bands and BPNN technique is applied to merge the results of the sub - bands. Sub -band technique can improve the time robustnesss of the threshold and BPNN can merge the output of the sub -bands non - linearly. The threshold trained with mixed data from the speaker himself and the imposters using ANN can improve the robustness against the indeterminacy of the imposters. The effectivity of the method presented here is testified by the experiments.