基于MAML和改进MobileNet模型的低龄儿童语音词汇识别  

Speech vocabulary recognition of young children based on MAML and improved MobileNet model

作　　者：马孟星 鄢元霞 马春晓 潘文林[3] MA Meng-xing;YAN Yuan-xia;MA Chun-xiao;PAN Wen-lin(College of Electrical and Information Engineering,Yunnan Minzu University,Kunming 650500,China;The Third Affiliated Hospital,Kunming Medical University,Kunming 650118,China;College of Mathematics and Computer science,Yunnan Minzu University,Kunming 650500,China)

机构地区：[1]云南民族大学电气信息工程学院,云南昆明650500 [2]昆明医科大学第三附属医院,云南昆明650118 [3]云南民族大学数学与计算机科学学院,云南昆明650500

出　　处：《云南民族大学学报（自然科学版）》2023年第3期346-351,共6页Journal of Yunnan Minzu University:Natural Sciences Edition

摘　　要：语言能力测试常用于评估低龄儿童的词汇储备等能力,运用语音识别等人工智能技术能够提高语言能力测试的工作效率,从而让更多的语言发育迟缓儿童能被尽早发现并得到治疗.低龄儿童语音相较于成人语音更难识别,且缺乏相关公开数据集,为了解决语言能力测试场景下的低龄儿童语音词汇识别问题,采集72名2~3岁儿童的语音数据,对具有参数少、计算成本低等特点的MobileNet模型进行了改进,并使用模型无关的元学习方法(MAML)优化改进模型,使改进模型适用于小样本环境下的低龄儿童语音词汇识别.实验证明,相关改进措施均能提高模型的儿童语音词汇识别性能.Language ability tests are often used to assess the vocabulary reserve and other abilities of young children.Artificial intelligence technologies such as speech recognition can improve the efficiency of the tests,so that more children with language development delays can be detected and treated as soon as possible.Comparedwith adult speech,young children's speech is more difficult to recognize,and there is a lack of relevant public data sets.In order to solve the problem of children speech vocabulary recognition in the scene of language ability tests,the speech data of 722-3-year-old children are collected from language ability tests.The MobileNet model with few parameters and low computational cost was improved,the improved model is suitable for few-shot learning of children speech vocabulary recognition by optimizing it using model-agnostic meta-learning(MAML).Experiments show that the relevant improvement measures can improve the model's performance of children's speech vocabulary recognition.

关 键 词：语言能力测试 语音词汇识别 小样本学习 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]