NMF-based Target Source Separation Using Deep Neural Network

Deep Probabilistic NMF Using Denoising Autoencoders

“…The last layer of the multi-layered NMF is decomposed by semi-supervised NMF, instead of NMF. In [16], they attempt to improve encoding vector estimation by using deep neural network to learn a mapping between data vectors and the corresponding encoding vectors. The method proposed by [17] addresses the non-linearity issue by embedding autoencoders to the NMF framework and by training the network and the factor matrices together.…”

Combining non-negative matrix factorization and deep neural networks for speech enhancement and automatic speech recognition

“…The activation coefficients are then used as input features of the DNN, instead of raw spectral coefficients as in [9] or the log spectrum in [2]. For each frame of noisy speech (at index position t), we build a large vector composed of the concatenation of the activation coefficients of speechĥ S,t and noisê h N,t vectors extracted on each frame on an analysis windows of width (2K + 1) frames centred on the t th frame.…”

“…However, the linear mapping assumption used in SNMF will fail when speech and noise overlap in the feature domain or share similar bases. Several SNMF-based approaches have already addressed this limitation, first by jointly training the noise and speech basis vectors in order to produce more discriminant subspaces [6,7,8], and also by using nonlinear mapping functions (typically with DNNs) to estimate the speech and noise coefficients [9].…”

“…Contrarily to [9], the DNN is here used to produce a non linear mapping function between the SNMF activation coefficients of noisy signals to the equivalent activation coefficients of clean speech. One motivation of pre-processing noisy recordings with supervised NMF is that the projection of noisy signals into the lower dimension of NMF may first reduce the complexity of DNN training, and also produce a better DNN initialization, thanks to injecting prior knowledge gained with unsupervised SNMF on training data.…”

“…Our proposal has been evaluated both for the tasks of Speech Enhancement and Automatic Speech Recognition (ASR) using several objective metrics. Our results have been systematically compared to several state-of-the-art DNN and SNMF-based SE systems [2,3,4,9]. Evaluations have been conducted using the framework provided recently by the CHiME-3 challenge [11] on speech separation and recognition on challenging real noisy speech recordings.…”