Conventional sub space based approaches for source localization use the spectral magnitude of MUSIC. In this paper, a group delay based method for source localization of spatially close speech sources over circular arrays, with minimal number of sensors is proposed. This approach is based on the MUSIC-Group delay spectrum and can be used to accurately estimate both azimuth and elevation angles of spatially close sources. Both simulated and real speech signal measurements are acquired over a circular array and the DOA estimation is carried out for several trials. The accuracy of the proposed approach is illustrated by using two dimensional scatter plots for a single source, and average error distribution plots for multiple sources. The high resolution property of this method is explained using the additive property of the MUSIC-Group delay spectrum. The proposed method is also evaluated under sensor perturbation errors. Experiments on distant speech recognition are conducted using the proposed approach on sentences from the TIMIT database acquired over circular arrays. The MUSIC-Group delay method indicates reasonable reduction in word error rates when compared to the standard MUSIC-Magnitude method as noted from these experiments.