Enhancing the Audio Achievement of the TIAGo Robot for Assistance in Medical Monitoring

Abstract Recently, there is an increased interest in the classification of acoustic events, and critical problems arise in real-life scenarios. Research has also provided a variety of robotic healthcare solutions to aid emergency response. This paper searches for a satisfactory audio signal identification algorithm for the TIAGo service robot to assist in medical applications. We found previously the difference between nine distinct classification algorithms which were applied to Moving Picture Experts Group 7, Mel Frequency Cepstral Coefficients, Linear Predictive Coding, and Linear Predictive Cepstral Coefficients features. Initially, all experiments were done on 1920 audio signals, using 10-fold cross-validation repeated 10 times to evaluate the accuracy of the classification. Using Linear Discriminant Analysis, k-Nearest Neighbors, and Support Vector Machine as classifiers, in conjunction with Mel Frequency Cepstral Coefficients as features, we have obtained the highest correct classification rates. Our new aim was to improve the accuracy upon the updated database, meaning 3300 audio signals. Using 34 and 64 Mel Frequency Cepstral Coefficients in the feature extraction phase we have obtained the highest value for correct classification rates for all experiments performed and small computation time. Six models were implemented for the final speech and indoor activities recognition algorithm. To improve the accuracy we applied a Grid search algorithm for the Support Vector Machine classifier. To get a satisfactory solution, 5-fold cross-validation repeated 20 times has been used.