Abstract
This thesis encompasses modern age feature extraction and machine learning tools for the detection of infant cry signals. In developmental psychology, infant crying is a measure of distress and an automated tool for measurement is extremely important. Initial testing is done using the MIRToolbox in MATLAB which proved to be non-ideal for real time signal analysis due to slow speed and poor memory management in MATLAB. Thus, a python based real time feature extraction tool Librosa is used to calculate parameters MFCC, delta-MFCC, pitch, zero-crossing, spectral centroid and energy of the signal. A large chunk of 21 minutes cry signal is used for feature extraction and used for the training of the crying segment. A similarly sized non-cry segment consisting of other sounds as speech, baby whimÂpering, toy sounds etc. are used to train the non-crying part of the model. A 111091 instances and 28 attributes based dataset (.CSV) is developed for our clasÂsifier. A Zero-R rule is used for baseline eastablishment, followed by classification using a random tree bagger and a multilayer perceptron. An error rate of 2.5% was achieved using 100 trees and 10 fold cross-validation using random forest and a similar result was achieved with a MLP with 500 training epochs, 15 hidden layers and 10 fold cross-validation. This is a major improvement over the existing results and the methods used for cry detection and much faster in feature extraction and is capable of handling much larger chunks of data for training and testing.