A Study Of Two Microphone Based Noise Suppression Algorithms For Bluetooth Headsets

The noise suppression algorithms for two microphone based Bluetooth headsets are investigated in this thesis. The relationship between noise reduction and speech distortion is analyzed and the effect of head on the acoustics path from source to microphones is studied. A realizable noise suppression algorithm for two microphone based Bluetooth headsets is realized. The performance is evaluated objectively.Firstly, this thesis categorizes various kinds of multi-channel noise suppression algorithms into two groups for analyzing two microphone based noise suppression algorithms for Bluetooth headset systems. One group is the coherence based algorithms, which employ the coherence function between signals in the two channels for noise suppression. The other group is the spatially preprocess based algorithms, which take advantage of spatial information to separate the noise reference signal from the noisy signal. An optimal solution with speech distortion constraint is analyzed. It needs to compromise between noise reduction and speech distortion. The head influences the high frequency response obviously, which enlarges the energy difference between two microphones.A realizable noise suppression algorithm based on the generalized sidelobe canceller (GSC) structure is proposed. This adaptive transfer function GSC (ATF-GSC) algorithm contains 4 parts:adaptive beamforming and block matrix, adaptive noise cancellation, voice active detector (VAD) and post-filtering. The block matrix decides the speech leakage degree. The VAD controls the update of each filter to ensure the stability. A post-filtering algorithm is used to further reduce the residual noise.Two typical algorithms are assessed based on the performance of noise reduction, speech distortion and overall performance. Different block matrix modeling methods and different mismatch situations of the ATF-GSC algorithm are evaluated. Evaluation results indicate that better overall performance can be achieved by taking balance between noise reduction and speech distortion.