|Supervisor:||Dr.-Ing. Klaus Reindl|
|Faculty:||Prof. Dr.-Ing. Walter Kellermann|
Microphones which should capture signals from distant sources will always contain noise and/or interfering signals. Therefore, for many applications as e.g., hearing aids, the acquired signals have to be „cleaned up“ in the sense that stationary and nonstationary noise and interfering speakers are suppressed, while simultaneously preserving the intelligibility of the desired source(s).
The problem of noise suppression is extensively studied in literature for single-channel as well as multichannel applications. Based on the cost function, there are different statistically optimum filters. However, all of them have in common that statistical information about the noise components (usually in terms of power spectral densities) is necessary to realize such concepts. Due to the nonstationary nature of speech signals, the second-order statistics of noise and interference components are also significantly (temporally) nonstationary. This makes it difficult to estimate PSDs reliably. Recently, a novel method for estimating (temporally) nonstationary PSDs has been proposed and is based on the spatial coherence which is assumed to be predominantly influenced by the geometric setup.
In this thesis, the (temporal) stationarity of the spatial coherence of noise and interference components should be analyzed. For this analysis different recordings for multispeaker and noisy scenarios have to be done in the lab. The stationarity of the PSDs itself as well as the stationarity of the spatial coherence for the different conditions has to be compared and investigated theoretically as well as experimentally by Matlab simulations. Well-documented and well-structured software is important. The thesis can be written in German or English.