透過您的圖書館登入
IP:3.144.199.9
  • 會議論文
  • OpenAccess

嬰兒哭聲偵測之特徵提取研究

摘要


哭是嬰兒向父母表達需求的一種方式,現今在許多育嬰產品中開始加入了哭聲偵測與辨識的技術。然而,在實際的應用場域中,這些設備所收集到的音訊不全然是嬰兒哭聲,還包含了許多周遭環境的聲音,因此本論文將提出濾除非嬰兒哭聲的技術。本研究比較了嬰兒哭聲偵測在不同分類器下的效能,提取了58個音訊特徵,並使用循序前進搜尋演算法與費雪鑑別率分別找出支援向量機、隨機森林、CART-Adaboost等三種不同分類器所適用的特徵集。實驗結果顯示支援向量機的準確率達92%,隨機森林91%,CART-Adaboost為89%。

並列摘要


Crying is the way that infants express their needs to their parents. Many baby products had provided the function of crying detection and recognition nowadays. In order to analyze the audio collected from the baby products, it is very critical to remove the noise that affects the result of the analysis. In this paper, we compared the capability of crying detection under different classifiers. A total of 58 acoustic features were extracted, then the Sequential Forward Selection algorithm and Fisher’s Discriminant Ratio are adopted to figure out the best feature set matched for each classifiers. Three classifiers, including Support Vector Machine, Random Forest and CART-Adaboost were used to compare the performance. The experiments show that the recognition accuracy of SVM, Random Forest, and AdaBoost are 92%, 91%, and 89%, respectively.

延伸閱讀