Typically, microphone signals acquired by human/human – interfaces (e.g. hearing aids, hands-free communication devices) and/or human/machine – interfaces are corrupted by stationary and/or nonstationary noise signals. Thus, communication via hands-free devices can be rather difficult and automatic speech recognizers cannot work reliably. Therefore, it is desirable to „clean up“ the microphone signals before they are stored, analyzed, or played out. This requires that the desired speech components are extracted from the noisy mixture signal by suppressing all interference and noise signals while the distortion of the desired speech components should be minimized.
This speech enhancement problem has been an active area of research since the invention of the spectral-subtraction technique in the middle 1960s. Currently, there are many different noise reduction techniques under investigation. Due to the steady increase in the processing power of digital signal processors (DSPs), multichannel techniques became recently a viable option.
In this thesis two recent multichannel noise reduction techniques should be investigated: Multichannel Wiener filtering and spatial prediction. These two methods should be compared and investigated for speech being corrupted by nonstationary noise. A theoretical as well as an experimental investigation by Matlab simulations is necessary. Well-documented and well-structured software is important.
Matlab, Digital Signal Processing course, interest in audio