Performance improvement of GSC algorithms by near channel subtraction-based blocking matrix

For generalized sidelobe canceller (GSC)-based multi-channel speech enhancement, reference noise generation is essential to reduce remaining noises after fixed beamforming. However, the target signal leakage in the blocking matrix of GSC is inevitable. Thus, in this paper, near-channel subtraction scheme is proposed to improve the blocking matrix of the GSC tightly related with the reference noise. The performance of the proposed algorithm is compared with that of the conventional reference channel subtraction algorithm in the aspect of speech recognition rates and output SNR gains.