Analysis on reliability of acoustic features for material recognition of damped impacted plates

Automatic recognition of the material of struck plates was investigated. Impact sound produced by three kinds of plate materials are synthesized using the sphere-plate collision model. We extracted lots of acoustic features in view of time-domain, frequency-domain and auditory perception respectively, and exploited a three layers BP neural network as the classifier. In order to test the stability of features in the present of noise, we added white noise to synthesized sounds and obtained the results of material recognition for different signal to noise rates. The results showed that effective duration, equivalent duration and decay constant derived from temporal envelopes of sounds perform well in both precision and robustness for material recognition.