Discriminative non-negative matrix factorization for single-channel speech separation

Non-negative matrix factorization (NMF) has emerged as a promising approach for single-channel speech separation. In this paper, we propose a new method of discriminative learning of NMF. In contrast to conventional approaches where the basis vectors are learned independently on clean signals from each speaker, our approach optimizes all basis vectors jointly to reconstruct both clean signals and mixed signals well. Our empirical studies validated our approach. Specifically, discriminative NMF outperforms standard methods by a large margin in improving signal-to-noise ratio for reconstructing signals.