Iterative blind source separation algorithm is often equivalent to a forward neural network trained by the unsupervised learning. Training iteration of parameters should be initialized beforehand. In this paper, an initialization algorithm is proposed for the blind separation of mixed speech signals based on the geometric structure of speech signal space. After the mixed signals are whitened, the quadrants of coordinates are regarded as the local PCA subspaces of the obtained signals. The mixing matrix can be estimated by the first eigenvectors of these subspaces. Simulation results show that separation performance of the FASTICA algorithm is improved by the proposed initialization algorithm.
展开▼
