Impulsive Timing Detection Based on Multi-Frame Phase Voting for Acoustic Event Detection

This paper proposes impulsive timing detection based on multi-frame phase voting for Acoustic Event Detection (AED). Since an impulsive sound exists for only a short time, the accuracy of the event timing plays an important role in the reduction of labeling costs and in the improvement of learning ef-ficiency. We propose here a method that detects impulsive sounds on the basis of acoustic signal processing using phase slopes, which are among characteristics peculiar to impulsive sounds. Phase slopes from multiple frames are converted to sample points and aggregated by weighted voting. Evaluation employing frame-wise F-scores under noisy environmental conditions shows an improvement of 0.22 over a conventional single frame method for speech processing.