The invention discloses a
frequency domain voice blind
separation method for multi-frequency-band switching call media node (CMN) nonlinear function, which belongs to the technical field of
speech enhancement and is characterized in that
frequency domain speech is divided into the two frequency bands of
low frequency and
middle frequency based on kurtosis distribution characters. Three types of multi-frequency-band schemes are applied for switching a nonlinear function of a plurality of CMN algorithms, at least one scheme is led to be most matched with
Gaussian performance and symmetry of
frequency domain speech. Compared with single nonlinear function CMN algorithms, the frequency domain voice blind
separation method for multi-frequency-band switching CMN nonlinear function is capable ofbeing adapted to voice changes in terms of
Gaussian performance and symmetry, and remarkably improves voice separation performance. When an ordinary amplitude
correlation method is adopted for voice sequence regulation of all frequency points, the separation
signal to
noise ratio of two paths of voice can be increased by 11dB at most, and the frequency domain voice blind
separation method is stable is performance, easy in
software and hardware achievement, and capable of being widely used in key technologies of computer
perception and decision-making, unmanned driving and the like so as to achieve the
speech enhancement function, and further improves entire performance of voice
signal processing tasks such as voice recognition and content understanding.