Learning to recognize horn and whistle sounds for humanoid robots

The efficiency and accuracy of several state-of-the-art algorithms for real-time sound classification on a NAO robot are evaluated, to determine how accurate they are at distinguishing horn and whistle sounds in both optimal conditions, and a noisy environment. Each approach uses a distinct combination of an audio analysis method and a machine learning algorithm, to recognize audio signals captured by NAO’s four microphones. A short summary of three audio analysis preprocessing methods is provided, as well as a description four machine learning techniques (Logistic Regression, Stochastic Gradient Descent, Support Vector Machine, and AdaBoost-SAMME) which could be used to train classifiers which would distinguish whistle and horn signals from background noise. Experimental results show that for each of the acquired data sets, there are multiple high-accuracy solutions available. Actually, the accuracy and precision results were all so high, that a more challenging dataset is needed to determination which method is optimal for this application.