Abstract
This thesis proposes a noise reduction system that takes advantage of far-field microphone array beamforming technology and combines with statistical-model based noise reduction algorithm in order to suppress noise and improve the signal-to-noise (SNR) ratio. This thesis also provides an effective solution for such situations that noises turn on and off as a step signal in real time implementation. For the front-end processing, the signal is converted to frequency domain by Fourier transform. This thesis adopts the property that the direction of each source is independent. Applying Minimum Power Distortionless Response (MPDR) beamformer, highly directional microphone array is designed to focus on the direction of main speech signals and suppress noise and interference from the background. After the beamforming processing, a single-channel signal is produced. Using Log Minimum Mean Square Error (Log-MMSE) estimator to estimate an optimize gain correction function as a post-filter, clean main speech signal can be obtained. This algorithm architecture can effectively reduce the interference on side-lobe and stationary noise on main-lobe. For evaluating the difference between enhanced signal and original signal, subjective listening test and objective Perceptual Evaluation of Speech Quality (PESQ), Segmental SNR (segSNR) test are applied. Automatic Speech Recognition (ASR) experiment shows the high performance in NR.