An Eigenvalue-based Method for Estimating the Number of Simultaneous Speakers

We address the problem of speaker number detection from speech signals of simultaneous speakers, collected by uniform linear microphone arrays (ULAs). Using the signal captured by microphone array a covariance matrix can be constructed. To take advantage of the information that covariance matrix provides, we employ eigenvalue decomposition on the covariance matrix. By finding the notable gap between decreasingly sorted eigenvalues of the covariance matrix, the number of the speakers can be determined. While many existent method for speaker number detection are faulty in the presence of the noise, our simulations on different numbers of simultaneous speakers demonstrates the robustness of the proposed method against both noise and reverberation