【作者】 闫杰； 朱晓莲； 谢磊； 李鹏； 梁家恩；

【Author】 Jie Yan;Xiaolian Zhu;Lei Xie;Peng Li;Jiaen Liang;School of Software and Microelectronics,Northwestern Polytechnical University;NPU-Unisound Joint Lab on Intelligent Speech Interaction;Beijing Unisound Information Technology Co.,Ltd;

【机构】 西北工业大学软件与微电子学院； 西北工业大学-云知声智能语音交互联合实验室； 北京云知声信息技术有限公司；

【摘要】 说话人年龄分类是语音处理中一个极具挑战性的任务,其关键在于如何提取具有判别性和鲁棒性的特征去构建分类器。近年来以DNN为代表的深度学习技术在语音识别等任务上的成功应用促使我们尝试使用神经网络去改进年龄分类系统。本文提出使用DNN提取Bottleneck特征去构建年龄分类系统的方法,同时结合Bottleneck特征和原始声学特征进行i-vector建模。在aGender语料库上的实验结果表明,该系统的总体分类正确率达到56.11%,比仅使用传统声学特征的系统相对提高18%。同时,随着BN特征维数的增加,整体分类正确率也会有进一步提高。更多还原

【Abstract】 Speaker age classification is a very challenging task in speech processing.One of the key factors of this task is to extract effective and robust features to build a decent classifier.In recent years,DNN has achieved great success in speech recognition,which motivate us to use neural network to improve the age classification performance.In this paper,we extract bottleneck features with DNN and build an i-vector system for age classification together with the traditional acoustic features.Experimental results show that the proposed approach achieves significant improvement in speaker age and gender classification.The overall classification accuracy of our system is 56.11%,which is 18% higher than that of the system using only traditional acoustic features.更多还原