A Morphological Approach to Single-Channel Wind-Noise Suppression

Today, a variety of technical devices deploy spoken language processing technology. In many practical use cases, not only stationary ambient noises but non-stationary interferences, such as wind noise, degrade their performance. If only one microphone is available, regular noise reduction schemes are capable of suppressing stationary interferences. Wind noises, however, are highly non-stationary and may be annoying in a hands-free application, for instance. To suppress such interferences in a speech signal, we propose to consider the signal spectrogram as an image and to exploit neighborhood relations in the time-frequency plane, hence morphological features. To this end, wind-noise-typical energy patterns are extracted by morphological operations employing both the signal spectrogram and its temporal derivative. The resulting estimate of the interference power spectral density is then used for speech enhancement in a subband-processing framework. Distance measures and informal listening tests indicate the effectiveness of the proposed method compared to classical noise-reduction schemes in the presence of non-stationary interferences of wind noise.