Abstract: Aimed at the features of speech signals which were non-stationary, unmodulated and had high fractional bandwidth, a novel direction-of-arrival (DOA) algorithm of coherent speech signals was proposed which was based on two array elements. The algorithm first selected the optimal frame which had higher energy and more uniformly distributed power spectrum at the range of low frequencies from the received signals of the two array elements, then divided it into the non-overlapping narrowband portions, afterwards used the Hilbert transform to convert them into analytic signals and made the different frequencies of the signals equivalent to each array elements of the virtual array, at last eliminated the coherence of the signals to obtain DOA estimation. A research of simulation experiment was given from multiple aspects, e.g. number of signals, sub-band division, signal-to noise ratio (SNR) etc and the effectiveness of the new algorithm was verified. From the simulation results and analyses, it is showed that the proposed algorithm has fine applicability, particularly when SNR is high. It is able to overcome the problem which two array elements can estimate only one signal source and achieve the DOA estimation of multiple coherent signals.