Development of automatic cough monitoring, measurement and service system

自动咳嗽监测、测量及服务系统开发

• 批准号: 20K12080
• 负责人: MARKOV K
• 金额: $ 2.75万
• 依托单位: The University of Aizu
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 2020
• 资助国家: 日本
• 起止时间: 2020-04-01 至 2024-03-31
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-20K12080/
• 关键词: cough recognition; cough monitoring; DNN audio analysis; machine learning; deep learning

During this year we built a high performance cough detection and monitoring system and successfully evaluated it using the data collected at FMU. There are 11 patients audio recordings in the dataset with highly irregular cough events time distribution. Out approach is to segment the input data into 10sec long segments and process each segment separately. Out model is based on fine tuned large audio model called HuBERT which can identify cough frames with high accuracy. In addition, we trained a special network which estimates the probability of each identified cough frame being the first, second, etc, in the couch event. This way, we can distinguish separate cough events even when they come right one after another within a long sequence of cough frames.

在这一年中，我们建立了一个高性能的咳嗽检测和监测系统，并成功地使用在FMU收集的数据进行了评估。数据集中有11例患者的音频记录，具有高度不规则的咳嗽事件时间分布。我们的方法是将输入数据分割成10秒长的段，并分别处理每个段。Out模型基于微调的大型音频模型HuBERT，可以高精度地识别咳嗽帧。此外，我们训练了一个特殊的网络，该网络估计每个识别出的咳嗽帧在沙发事件中是第一个、第二个等的概率。这样，我们可以区分单独的咳嗽事件，即使它们在一长串咳嗽帧中一个接一个地出现。