The invention belongs to the technical field of digital signal processing and discloses a voice underdetermined blind identification method and device based on frequency spectrum correction and data density clustering to improve algorithm precision, improve algorithm efficiency remarkably and improve the robustness of the algorithm on noise. According to the technical scheme, the method comprises the steps of 1, conducting hanning windowed L point 50% overlapped short time Fourier transform (STFT) to obtain an observation frequency spectrum Xm(t,k); 2, conducting frequency spectrum correction on the STFT observation frequency spectrum frame by frame; 3, for a specific time frame t0, conducting mode purification on all harmonic wave parameters; 4, executing the step 2 and the step 3 frame by frame, and collecting SAS modes of the all the time frames to form a monosource domain omega={zi, i=1,...,P}, wherein P is the number of modes of the monosource domain; 5, conducting data density clustering on the SAS modes in the monosource domain. The method and device are mainly applied to digital signal processing.