We propose a time-domain blind source separation (BSS) algorithm that utilizes geometric information such as sensor positions and assumed locations of sources. The algorithm tackles the problem of convolved mixtures by explicitly exploiting the non-stationarity of the acoustic sources. The learning rule is based on second-order statistics and is derived by natural gradient minimization. The proposed initialization of the algorithm is based on the null beamforming principle. This method leads to improved separation performance, and the algorithm is able to estimate long unmixing FIR filters in the time domain due to the geometric initialization. We also propose a post-filtering method for dewhitening which is based on the scaling technique in frequency-domain BSS. The validity of the proposed method is shown by computer simulations. Our experimental results confirm that the algorithm is capable of separating real-world speech mixtures and can be applied to short learning data sets down to a few seconds. Our results also confirm that the proposed dewhitening post-filtering method maintains the spectral content of the original speech in the separated output.