Automatic classification of audio through fruit percussion: toward no destructive estimation of quality

In this paper, we present the prototype of a system designed to classify different fruits based on the sound generated by the avocado when is percussion it. The advances andresults shown in this article are part of the development of a computational system for the estimation of avocado quality parameters from the analysis of the sound resulting from percussion. We sought to identify key elements such as the components that characterize the audioof each fruit; different audio signal processing and characterization techniques and different ANN (Artificial Neural Networks) architectures were tested to build an automatic classifier. For the training and testing stages, a total of 270 audios resulting from the percussionof tomatoes, onions, and avocados with 90 samples of each fruit were used. The results of the different tests show how the PM (Multilayer Perceptron) turns out to be the best architecture for the classification model using the Log-Mel spectrogram to characterize the signal. With this combination an average of 96% accuracy was achieved, demonstrating that by using machine learning it is possible to accurately classify the sound of different fruits.