基于预训练模型的多示例语音内容分类  

Multiple Instance Learning for Speech Content Classification with Pre-trained Model

作　　者：张建宜 姚佳奇 褚衍杰 燕继坤 梁杰 ZHANG Jianyi;YAO Jiaqi;CHU Yanjie;YAN Jikun;LIANG Jie(Information Engineering University,Zhengzhou 450001,China;National Key Laboratory of Science and Technology on Blind Signal Processing,Chengdu 610041,China)

机构地区：[1]信息工程大学,河南郑州450001 [2]盲信号处理国家级重点实验室,四川成都610041

出　　处：《信息工程大学学报》2022年第2期141-147,共7页Journal of Information Engineering University

基　　金：国家自然科学基金资助项目(61272041)。

摘　　要：语音内容分类主要用于对大批量信号进行自动处理,并基于用户的兴趣选择语音文件。据此提出了一种新的分类方法,在多示例学习框架下,使用无监督语音表示学习对大规模未标记数据进行预训练,得到用于提取语音深层表示的预训练模型,提取的语音表示作为下游分类器的输入。真实语音数据集上的实验结果表明,多示例学习在处理语音分类问题上具有优势,提出的方法能够提高分类的效果,在平均准确率指标上优于3种基线方法。Speech content classification is mainly used to automatically process signals and select files based on user interest.According to this,a new method is proposed.Under the multiple instance learning framework,unsupervised speech representation learning is utilized to pre-train a model for extracting deep speech representation as the input of downstream classifier with large-scale unlabeled data.Experimental results on real-world speech dataset indicate that multiple instance learning has an advantage in speech classification,and the proposed method can improve the classification performance and is superior to the three baseline methods in average precision metric.

关 键 词：语音内容分类 多示例学习 预训练模型 无监督语音表示学习 

分 类 号：TN912.3[电子电信—通信与信息系统]