Pattern recognition system for automatic identification of acoustic sources

An intelligent recognition system was designed using pattern recognition techniques to distinguish the noise signatures of five different types of acoustic sources. Information for classification was calculated from the power spectral density and autocorrelation taken from the output of a single microphone. The system included a training step where it learned to distinguish the sources and automatically select descriptive quantities for optimal classification performance. Information learned in training was stored and used to identify recordings of the five types of sources presented during testing. Results of testing indicate that the current optimal design could correctly identify 90% of the recordings. Identification of noise corrupted signatures and identification of recordings not used in training is discussed.