Blind source separation based on NNAE-GCC

The performance of blind source separation algorithm based on non-negative matrix factorization is good, but the complexity of algorithm limits to on-line separation. We propose an acoustic separation method based on Non-Negative Automatic Encoder (NNAE) and Generalized Cross Correlation (GCC), which reduces the complexity of the algorithm and provides the possibility of online separation. The time-frequency feature of speech is mapped to high-dimensional space via NNAE. The time delay of sources is calculated by GCC. According to the time delay information, the delay masking vector of each frame of speech is calculated in high-dimensional space. The NNAE-GCC combines high-dimensional feature and spatial position information to implement blind source separation. The complementary spectral and spatial information can be simultaneously exploited to improve the performance of speech separation. Experiments on simulation mixed speech and the real-world data, which is collected from a circular microphone array with a diameter of 4 cm, show the potential of the proposed algorithm for speech separation in reverberant environments. By replacing the NMF with NNAE, we found that the neural network is given a clear physical meaning.