Underdetermined blind source separation (BSS) is discussed.Based on time-frequency analysis

an underdetermined BSS method is developed

in which sources are not necessarily sparse or independent mutually.First

we compute a sequence of matrices of time-frequency distributions (TFDs) and obtain the auto-source TF points

then we fold the TFD matrices into a third-order tensor and calculate the mixing matrix by tensor canonical decomposition

finally we obtain the sources by calculating the pseudo-inverse matrix and TF synthesis techniques.The simulations demonstrate that the proposed method outperforms the existing methods in performance significantly.