Noise reduction in multichannel medium

In this thesis, using a developed VAD algorithm; the efficiency of multichannel noise reduction algorithms such as Wiener filter, Speech Distortion Weighted Wiener Filter, Spatial Prediction Filter, and Minimum Variance Distortionless Response (MVDR) was compared. While four different audio signals are used in the algorithms, three different types of noise are added to these audio signals at different angles. Before calculating the filter coefficients, noisy speech and only the detection of noisy areas were performed using a voice activity detection algorithm. To be able to distinguish between noisy speech and only noisy areas in this algorithm; Short Time Energy, periodicity and spectral flatness are used. After noisy speech and noisy areas were determined, the coefficients of the filter were calculated with the algorithms mentioned above. Finally, the calculated filter coefficients and the frequency components of the reference microphone at the input are multiplied to obtain improved signals for each algorithm. Segmental SNR is an objective method for calculating the performance of algorithms and MOS is used as a subjective method. Although the performance of the SDW wiener filter is better in terms of noise reduction, better results are obtained with SP and MVDR filters in terms of speech intelligibility and perceptual quality.