ระบบนิวโรฟัซซีเพื่อจำแนกเสียงร้องของเด็กทารก

Abstract

This thesis presents recognition and classification algorithm for infant sounds. This algorithm has the highest accuracy comparing with other algorithms. The infant sounds used in this thesis are identified from an expert and there are 5 meanings including feeling hungry, feeling sleepy, feeling discomfort, having lower gas and needing to be burped. 251 data sets of infant sounds are tested in this thesis. The first process of algorithm used the popular feature extraction algorithm, i.e., Mel Frequency Cepatral Coefficients (MFCC), Perceptual Linear Prediction (PLP) and Relative Spectral (RASTA) for extract feature vectors from data set. Then, the data set is used in recognition and classification process. This process is divided into two parts. The first part, recognition, has three sub processes which are Neural Network, Data Normalization and Fuzzy Logic. The second part, classification, uses K-nearest Neighbor classifier. This thesis finds the appropriate structure from many experiments. The result shows that the proposed algorithm has higher accuracy in infant sound classification than other algorithms and has 86.25% accuracy. Moreover, this thesis designs experiments for using proposed algorithm with Thai words. In the experiment, there are many factors such as noise, pronunciation and number of words. The experimental result shows that the proposed algorithm has high accuracy in classification.