Research On Infant Speech Detection Algorithm Base On Machine Learning

In real life,crying is the main way for babies to communicate with the outside world.Infants communicate their needs to the outside world through cries,and infant caregivers are supposed to meet babies' needs based on their cries.At present,most of the research on baby-crying focuses on the classification of baby-crying and detecting the causes of baby-crying,such as hunger,drowsiness,and discomfort.However,the application of the above research results needs to be based on the precisely detection of baby-crying,which is seldom investigated at home and abroad.Therefore,this thesis mainly studies the core algorithms of baby-crying real-time detection system that can be applied to different scenes: speech endpoint detection,speech enhancement and baby-crying classification algorithm.Moreover,this thesis applies speech endpoint detection and speech enhancement algorithm to baby-crying real-time detection system for the first time.Specifically,the main work of this article is as follows:Firstly,a fuzzy C means(FCM,Fuzzy C-means)based baby-crying endpoint detection algorithm is proposed in this thesis.With the features of baby-crying,this algorithm innovatively applies the FCM clustering algorithm to the endpoint detection of baby-crying,which avoids the non-applicability of the FCM clustering algorithm in endpoint speech detection because a clustering center needs to be set in advance in the clustering algorithm.The simulation results show FCM-based baby-crying endpoint detection algorithm achieve higher accuracy than Xunbo's in 2018.Secondly,a baby-crying database is built in this thesis.Given that most data base used in baby-crying research are private,samples of baby-crying are collected through internet in order to satisfy the needs of the current research.The data sets after preprocessing which removes the unqualified interferences combine with noises from Noise92,forming baby-crying samples with different Signal to Noise Ratio.Thirdly,the thesis improves the autocorrelation speech enhancement algorithm.The simulation results indicate that the improved algorithm significantly reduces the residual music noise.In addition,the thesis introduces another four commonly used speech enhancement algorithm and gives their simulation results based on which the advantages and disadvantages of each algorithm are discussed.Fourthly,A baby-crying classification algorithm based on support vector machine(SVM)is proposed in this thesis.Compared with the deep-learning based baby-crying detection algorithm,the former is more applicable to the small-to-medium-sized data sets in the scenes that require real-time performance.Based on the features of baby-crying,this thesis extracts such features of the input speech signals as the pitch period,the spectral roll-off point,the sub-band frequency variance,the Mel frequency cepstral spectral coefficient,and the sub-band energy variance.The simulation results show that the accuracy of the SVM classification algorithm reaches 94%.Lastly,a real-time baby-crying detection system is built in this thesis by applying the baby-crying endpoint detection algorithm,speech enhancement algorithm as well as the baby-crying detection algorithm to the real-time baby-crying detection system.Specifically,this thesis implements a real-time baby-crying detection system which supports real-time baby-crying detection in any scene for more babies.Furthermore,a browser interactive page is built for users to catch the baby status in real time.