Application of non-negative spectrogram decomposition with sparsity constraints to single-channel speech enhancement

• We successfully separate speech from noise without any pre-trained models.
• We focus on difference in spectral distributions between speech and noise.
• Selectively imposing sparsity constraint on basis vectors allows no training or post-processing.
• Experiments on various real-world noises show the proposed method results in improved performance.

We propose an algorithm for single-channel speech enhancement that requires no pre-trained models – neither speech nor noise models – using non-negative spectrogram decomposition with sparsity constraints. To this end, before staring the EM algorithm for spectrogram decomposition, we divide the spectral basis vectors into two disjoint groups – speech and noise groups – and impose sparsity constraints only on those in the speech group as we update the parameters. After the EM algorithm converges, the proposed algorithm successfully separates speech from noise, and no post-processing is required for speech reconstruction. Experiments with various types of real-world noises show that the proposed algorithm achieves performance significantly better than other classical algorithms or comparable to the spectrogram decomposition method using pre-trained noise models.