Beam forming voice separation method based on complex space angle center Gaussian hybrid clustering model and bidirectional long and short time memory network

The invention discloses a beam forming voice separation method based on a complex space angle center Gaussian hybrid clustering model and a bidirectional long and short time memory network. In the training stage, a logarithmic power spectrum of a reference channel training voice signal is calculated, and sine and cosine values of a phase difference between the reference channel training voice signal and other channel training voice signals are used as input characteristics of the bidirectional long-short time memory network. Based on a complex space angle center Gaussian hybrid clustering model, calculating a masking value of each target sound source as a training target of the bidirectional long-short term memory network, and using mean square error loss as a loss function. In the test stage, according to the masking estimation value of each target sound source output by the bidirectional long-short-term memory network in the reference channel test voice signal, the covariance matrix of the multi-channel tes