Convolutive source separation and selective null beamforming using a pilot_based channel estimation technique

Abstract

In this paper, we propose new convolutive source separation and selective null beamforming method using a pilot-based channel estimation technique in reverberation environments. First, we show convolutive sound source separation method in determined and overdetermined cases. For the acoustic channel estimation, we propose new channel estimation method using pilot sequence conceptually similar to a pilot code of wireless communication. Acoustic channel estimation considered the effect of spectral sampling with frequencies matching STFT frequencies provides precise channel information we want to know and has the less computational loads than other competitive methods. After channel estimation, we can separate the signals from observations using estimated channel information. We show the proposed method has better performance than Griffith-Jim beamformer and Lee``s method in a various measures. Second, we show the new convolutive sound source separation method for underdetermined case. This method combines frequency domain beamforming and independent component analysis(ICA). The convergence and the result of the separation of gradient-based algorithms is very much influenced bye the initial value. This is a key motivation of proposed method. To use standard time domain ICA, the number of sources must equals the number of observations. We generated new input initial signals for ICA stage using frequency domain beamforming as same as the number of source signals. This approach provides well conditioned initial input signals for ICA stage and we can improve the performance of separation in underdetermined case. Finally, we propose selective null beamforming using estimated channel information. If we use same pilot sequence at each source position, we can obtain relative difference of acoustic channel between transmitters as well as between microphone array. From above channel information, we can calculate beamforming weight vectors to null the signals at target locat...