With the technological progress, devices, such as mobile phones, tablet computers or hearing aids, can be used in a large variety of every-day situations for mobile communication. Acoustic background noise signals, which are picked up with the desired speech signal, can impair the signal quality and the intelligibility of a conversation. A special noise type is generated outdoors, if the microphone is exposed to a wind stream resulting in strong-rumbling noise, which is highly non-stationary. As a result, conventional approaches for noise reduction fail in the case of noise induced by wind turbulences.


This thesis is focused on the development of signal processing concepts, which reduce the undesired effects of wind noise. The key contributions are:


• Signal analysis of wind noise


• Digital signal model for wind noise generation


• Signal processing algorithms for detection and reduction of wind noise signals.


All these topics are considered with the focus on the development of algorithms for single and dual microphone systems.


The analysis of recorded wind signals is the first step and gives valuable information for the estimation and reduction of wind noise. Furthermore it leads to a signal model for the generation of reproducible artificial wind noise signals.


For the enhancement of the disturbed speech, an estimate of the underlying wind noise signal is required. In contrast to state-of-the-art noise estimation algorithms, the spectral shape and energy distribution is exploited for the distinction between speech and wind noise components leading to a novel estimation scheme of the wind noise short-term power spectrum. Considering a system with two microphone inputs, the complex coherence function of the two recorded signals is exploited for wind noise estimation. In addition to commonly used noise reduction schemes by spectral weighting, an innovative concept for speech enhancement is developed by using techniques known from artificial bandwidth extension. Highly disturbed speech parts are replaced by corresponding parts from an artificial speech signal.


Objective measures indicate a significant increase of both the signal-to-noise ratio and the speech intelligibility. Besides, two application examples show that the proposed methods are very efficient and robust in realistic scenarios.