Abstract

For many years, enhancement of noisy speech signals has been an extensively researched topic. Many approaches have already been proposed which proved to be more or less effective. One aspect which certainly needs attention is the complexity of computation. A robust and simple enhancement method is more desirable for real time eplication. This thesis focuses on implementing and evaluating a pre-training based method which simplifies the computation of enhancement by taking advantage of the properties of an Air-Bone microphone. This device combines speech recording via air and via the skull of the speaker that should be enhanced. In contrast to statistically based methods, our method is simply analysing similarity between training and input data. The results show that a significant enhancement of a speech signal with heavy noise interference is possible. Nevertheless, few weak spots of the process have been uncovered, which motivate further development and refinement of the Air-Bone based approach for speech enhancement.