Signature whistle is individually distinctive vocalization that plays a vital role in the context of social interactions and group cohesion. However, sometimes the signals overlap recorded by hydrophones when several dolphins squawk simultaneously. This impedes the further research of signature whistle in vocal interactions. An efficient blind source separation technique that can identify and separate mixed signals is necessary. Here the joint approximate diagonalization of eigenmatrix (JADE) method is adopted to separate mixed signals in each frequency bin. To address the permutation problem induced by this algorithm, the spectral consistency of signature whistle is exploited. In simulation, we take the manually mixed dolphin whistles, which have been distorted by underwater acoustic channels, for separation. Results show that the original sources are effectively separated and the signal-tonoise ratio (SNR) is significantly improved.