论文部分内容阅读
在网络应用环境下,需要处理的音频数据和注册说话人急剧增加,传统说话人辨识方法难以满足实时性要求.文中提出采用K-L散度的说话人模型聚类方法,从而构造一个分级辨识模型,提高辨识效率.研究利用类辨识信息估计置信度的方法,可尽早有效排除集外说话人.实验结果显示,文中方法可使辨识速度平均提高3.2倍,而闭集辨识错误率平均只有0.9%的增加.采用类辨识置信度进一步提高开集辨识速度,并且在保持集内错误率不变的情况下,使集外错误率相对下降5.1%.
In the network application environment, the audio data and the registered speaker who need to be processed are increased sharply, and the traditional speaker recognition method can not meet the real-time requirement.In this paper, we propose a speaker classification clustering method using KL divergence to construct a hierarchical identification model, Improve recognition efficiency.Experimental results show that the proposed method can increase the recognition speed by an average of 3.2 times and the recognition rate of closed-set recognition average only 0.9% The class recognition confidence is used to further improve the speed of open set recognition, and the error rate outside the set is decreased by 5.1% while keeping the error rate unchanged.