Efficient blind speech separation suitable for embedded devices

A blind speech separation method with a low computational complexity is proposed. This method consists of a combination of independent component analysis with frequency band selection, and a frame-wise spectral softmask method based on an inter-channel power ratio of tentative separated signals in the frequency domain. The softmask cancels the transfer function between sources and separated signals. A theoretical analysis is given. Performance and effectiveness are evaluated via source separation simulations and a computational estimate, and experimental results show the significantly improved performance of the proposed method. The segmental signal-to-noise ratio achieves 7 [dB] and 3 [dB], and the cepstral distortion achieves 1 [dB] and 2.5 [dB], in anechoic and reverberant conditions, respectively. Moreover, there can be a reduction of over 80% in computational complexity compared with unmodified FDICA.