深度卷积和递归神经网络在咳嗽声音信号识别中的作用及意义

Cough frequency monitoring is a gold standard for objective assessment of cough severity and presents with a wide clinical application prospect. However, the current various methods for cough detection have not met the application requirements because of an insufficient recognized accuracy. Our previous studies have demonstrated deep neural network is superior to the current recognition models, but how to establish and optimize the detection model according to the inherent special features of cough signals and the cough’s temporal structure remains to be solved. Considering the advantages of deep convolutional and recurrent neural network on speech recognition and the similarity between the cough sound detection and speech recognition, we speculate that the use of the deep neural network in the model building for the discrimination of cough sound signals may capture the acoustic feature extraction reflecting the inherent special features of cough signals and establish the detection model fitting the cough’s temporal structure, and finally reach a more precise accuracy than the current detection models. In this proposal, we plan to extract inherent acoustic feature suitable for cough detection based on statistical characteristics of cough sound signals through convolutional neural network, then to establish the model for cough detection fitting the long term and short term temporal structure of cough signals by recurrent neural network, investigate its effect and implication, and validate the model in the large data of cough sound signals from the patients with cough, in order to improve the performance of cough frequency monitoring system and promote its clinical utilization.

咳嗽频率监测是咳嗽严重度客观评价的金标准，临床应用前景广泛。但现有的咳嗽监测技术咳嗽识别精度低，难于满足使用要求。我们的前期研究证实深度神经网络在咳嗽音的识别中优于现有识别模型，但仍未解决根据咳嗽声音信号内在特点和包含的时序结构进行特征优化建模的问题。鉴于深度卷积和递归神经网络在语音识别中的优越性以及咳嗽识别与语音识别之间的相似性，我们预计将其用于咳嗽声音信号识别建模有可能建立体现咳嗽声内在特点的咳嗽特征提取方法和能拟合咳嗽信号整体时序结构的咳嗽分类新算法，取得比现有识别模型精度更高的咳嗽识别率。本课题拟基于咳嗽信号内在统计特性，以卷积神经网络提取适于咳嗽识别的声学特征，契合咳嗽信号长时、短时时序结构，建立基于递归神经网络的咳嗽识别模型，并阐明其作用和意义，在大量临床采集的患者咳嗽声音信号上进行验证，以提高咳嗽频率监测的性能，推进其临床实用化。

咳嗽频率监测是咳嗽严重度评价的金标准，临床应用前景广泛。但现有的咳嗽监测技术咳嗽识别精度低，难于满足使用要求。鉴于神经网络在语音识别中的优越性以及咳嗽识别与语音识别之间的相似性，我们预计将其用于咳嗽声音信号识别建模有可能建立体现咳嗽声内在特点的咳嗽特征提取方法和能拟合咳嗽信号整体时序结构的咳嗽分类新算法，取得比现有识别模型精度更高的咳嗽识别率。为此，我们采集、标记与整理不同病因慢性咳嗽患者的咳嗽声音数据，基于卷积神经网络提取其咳嗽声学特征，在此基础上基于递归神经网络进行咳嗽检测器建模，随后通过对以上数据收集分析，建立基于嵌入式系统的便携式咳嗽声识别系统，并在临床上对该系统进行初步验证，以阐明其作用和意义。.研究结果显示1）针对咳嗽识别中的特征提取，使用一种更符合人耳听觉生理特性的特征组合，即Gammatone倒谱系数（GTCC），该特征通过一组人耳听觉特性的频率刻度（即ERB刻度）缩放频谱，之后使用Gammatone滤波器组进行滤波，进一步通过倒谱系数提取声学特征，相比常用的梅尔倒谱系数（MFCC）特征，提升了咳嗽识别性能；2）采用循环神经网络RNN传递不同时刻间的信息，以深度双向长短时记忆网络（Deep BiLSTM）进行前后时序信息处理，以批标准化优化Deep BiLSTM网络训练过程，通过ADADELTA算法进行参数更新，最终能完成基于嵌入式系统的深度神经网络便携式咳嗽声识别系统设计；3）在慢性咳嗽患者上的初步验证研究显示该模型的咳嗽信号识别对个体患者有很好的敏感性和精度，能满足一般临床使用，但尚需通过模型进一步优化和更多数据的深度学习进一步完善。.研究证实Gammatone倒谱系数在咳嗽特征提取中优于梅尔倒谱系数，循环神经网络联合深度双向长短时记忆网络能有效进行咳嗽音信息处理，基于嵌入式系统的深度神经网络便携式咳嗽声识别系统为咳嗽检测的临床实用化创造了条件。