Single channel source separation using time–frequency non-negative matrix factorization and sigmoid base normalization deep neural networks

Abstract

Conventional single channel speech separation has two long-standing issues. The first issue, over-smoothing, is addressed, and estimated signals are used to expand the training data set. Second, DNN generates prior knowledge to address the problem of incomplete separation and mitigate speech distortion. To overcome all current issues, we suggest employing single-channel source separation with time–frequency non-negative matrix factorization, as well as sigmoid-based normalization deep neural networks. The proposed system consists of the two steps listed below. The first is the training phase, and the second is the testing phase. The difference between these two testing and training stages is that the testing stage uses a single-channel multi-talker input signal and the training stage uses a single-channel clean input signal. Both of these testing and training stages send their input signals to Short Term Fourier Transform (STFT). STFT converts input clean signal into spectrograms then uses a feature extraction technique called TFNMF to extract features from spectrograms. After extracting the features, using the SNDNN classification algorithm, and the classified features are converted to softmax. ISTFT then applies to softmax and correctly separates speech signals. Investigational outcomes demonstrate that the proposed structure attains the best result associated with accessible practices.