Mangosteen export generates large revenue; however, translucent mangosteens, which contain undesirable internal condition, result in the shipment rejection and decrease the reliability of the export. This research investigates a novel non-destructive classification approach based on acoustic frequency response to detect mangosteens containing translucent fleshes. The set of uniform-distributed multi-frequency acoustic signal is generated and passed through each mangosteen under the test. The frequency responses, describing a feature space, for all mangosteens are computed via the discrete Fourier transform. To prevent intensive computation, a linear optimization is adopted to select relevant frequency contents, creating a compact classifying feature vector. To solve the classification problem, two proposed acoustic-based classification techniques are studied, namely linear classifier (LC), and non-linear classifier (NLC) based on an artificial neural network. Then the results from both classifiers are compared against the results from the conventional water-floating (WF) approach. Against the experimental data, it is found that the average flesh classification accuracy of good mangoteens achieved from the LC and the NLC are about 61% and 74% respectively, while the WF yields an accuracy of about 69%. It is evident that the acoustic-based approach possesses the convincing accuracy for solving the problem of export-grade translucent mangosteen classification. In addition, the paper shows that a mangosteen's physical density can possibly provide intuitive information for better classification performance in the future research study. © 2012 IEEE.