DNN-Based Frequency Component Prediction for Frequency-Domain Audio Source Separation

Multichannel audio source separation (MASS) plays an important role in various audio applications. Frequency-domain MASS algorithms such as multichannel nonnegative matrix factorization achieve better separation quality. However, they require a considerable computational cost for estimating the frequency-wise separation filter. To solve this problem, we propose a new framework combining the MASS algorithms and a simple deep neural network (DNN). In the proposed framework, frequency-domain MASS is performed only in narrowband frequency bins. Then, DNN predicts the separated source components in other frequency bins, where both the observed mixture of all frequency bins and the separated narrowband source components are used as DNN inputs. Our experimental results show the validity of the proposed MASS framework in terms of computational efficiency.