论文部分内容阅读
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。
This paper presents a robust speaker verification threshold design method based on subband technology and artificial neural network technology. The effective frequency band of speech signal is divided into several subbands to train or recognize independently. Output data fusion based on the final verdict. Each sub-band model training and identification using vector quantization technology, data fusion technology using BP artificial neural network. Adopting the subband technology can improve the time robustness of the speaker to confirm the threshold. The neural network technique is used for the non-linear data fusion on the output of each subband on the one hand, and the data and the fake of the interlocutor on the other hand Participants’ data is mixed training so that the training thresholds identified are robust to presumptive uncertainty. The design method proposed in this paper can get a robust threshold of recognition and has been experimentally verified.