Impact-noise suppression with phase-based detection

Abstract

This paper proposes impact noise suppression with a new phase-based detection. Different from any other conventional algorithms which rely on magnitude-based or model-based detection, a new impact-noise detection utilizing phase linearity of the input noisy signal is developed based on an analysis of an impulse. Phase slopes of the input noisy signal is compared with an ideal phase slope obtained from the peak magnitude of the time-domain noisy signal. Phase unwrapping problem is alleviated by the use of a rotation vector of frequency domain components. Evaluation with PESQ for push-button clicks shows an improvement of 0.9 over a conventional noise suppressor for voice communication. Comparison of enhanced signal spectrogram with that of clean speech demonstrates superior enhanced signal quality. The proposed phase-based detection may be combined with an amplitude-based detection for more accurate results.