Discrimination of blasting sounds, music, and wind noise using a Gaussian mixture‐model classifier.

Abstract

Wind noise and acoustic signals were recorded outdoors with a horizontally oriented, square, 7×7 microphone array having an intersensor spacing of 0.914 m. Even with this very small spacing, the wind noise was found to have very little spatial correlation at frequencies above 15 Hz, thus indicating that the turbulent eddies responsible for the wind noise distort very rapidly. This property can be used to distinguish wind noise from acoustic signals possessing similar frequency content. To demonstrate this idea, wavelet processing was applied to recordings with predominantly wind noise, propane cannon blasts, and unsteady but persistent sounds (music). A set of features involving ratios of the wavelet coefficients at different frequencies and between different microphones was then formulated. Three of the features relate to the time‐domain shape of the signals, two to the spectral content, and three to the spatial correlation. A Gaussian‐mixture‐model classifier was trained with the feature statistics of the three signal types, and classifier predictions were then compared against an independent test data set. Results indicate a 96.6% correct classification rate of the signals. Signal shape features reliably distinguish the blasting and music, whereas the coherence features reliably distinguish the wind noise from acoustic signals.