Modular NMF and DNN Speech Enhancement Approach with Update Noise Base

This paper presents a novel modular speech enhancement technique consisting of statistical models, namely non-negative matrix factorization (NMF) and Minimum-Mean-Square-Error (MMSE) with a deep neural network (DNN) as a supervised algorithm. Here, we propose improving the NMF results, especially in the presence of unseen non-stationary noise with an online-update-noise-basis procedure (ouNMF). This algorithm extracts noise basis in real experiments from the noise-only segments detected by a voice-activity-detection (VAD), and update it appropriately in a regularization process. Moreover, we propose a deep neural network (DNN) module to further enhance speech activation coefficients of ouNMF to reduce the noise residue present at the output of the previous module. We have shown that an MMSE module along with the proposed approach can compensate for the drawbacks of supervised DNN in case of unseen non-stationary noises. The experimental results performed on TIMIT database showed that the proposed approach outperforms the baselines including statistical and NMF -based approaches in terms of perceptual evaluation of speech quality (PESQ).