Speaker
Description
Human hearing can reliably distinguish materials based on impact sounds, and machine learning methods can replicate this ability when acoustic differences are pronounced. However, discriminating between materials that sound highly similar even to the human ear remains challenging.In this study, we demonstrate that glass and metal can be effectively separated using simple thresholding of acoustic features. In contrast, classifying eight types of optically identical plastic parts of identical geometry proves significantly more difficult. For this eight-class task, models trained on 48 kHz signals achieve an accuracy of 25%, exceeding the random baseline of 12.5%. Increasing the sampling rate to 192kHz improves performance to 36%, while cascading classification strategies further raise accuracy to 47%, with some plastic types exceeding 60% class-wise accuracy.These results show that acoustically distinguishing visually identical plastic materials is feasible with moderate accuracy. The experimental setup ensures that observed signal variations arise solely from intrinsic material properties.