Direction-of-arrival (DOA) estimation has been one of the important issues in acoustic signal processing. It is well known that the DOA can be estimated based on the cross correlation with stereo observations. It is, however, difficult to accurately estimate the DOA because the cross correlation-based function could not yield the sharp spatial feature. On the other hand, the multiple signal classification (MUSIC) algorithm provides the pointed spatial feature but it requires the exact number of sound sources in advance. In this paper, the cross correlation-based spatial feature is sharpened using a neural network. The neural network is trained using the cross correlation-based spatial feature calculated from the stereo observations and the binary teacher signals. It is confirmed that the non-linear cross correlation-based spatial feature has an extremely sharp peak at the direction of the target sound source without the information on the number of sound sources.
