A time-frequency preprocessing method for blind source separation of speech signal with temporal structure

Determination of the number of sources is a practical issue that has to be addressed in applications of underdetermined blind source separation (UBSS). This paper proposes a noise-robust UBSS algorithm for highly overlapped speech sources in the short-time Fourier transform (STFT) domain. The basic principle of the proposed algorithm firstly estimates the unknown number of sources in time-frequency domain. Secondly, the original sources are recovered by a separation method using both sparseness and temporal structure. To mitigate the noise effect on the detection of auto-source TF points, we propose a method to effectively detect the auto-term locations of the sources by using the principal component analysis (PCA) of the STFTs of noisy mixtures.