15th Annual Conference of the International Speech Communication Association

September 14-18, 2014

NMF-Based Speech Enhancement Incorporating Deep Neural Network

Tae Gyoon Kang (1), Kisoo Kwon (1), Jong Won Shin (2), Nam Soo Kim (1)

(1) Seoul National University, Korea
(2) GIST, Korea

Recently, lots of algorithms using machine learning approaches have been proposed in the speech enhancement area. One of the most well-known approaches is the non-negative matrix factorization (NMF) -based one which analyzes noisy speech with speech and noise bases. However, NMF-based algorithms have difficulties in estimating speech and noise encoding vectors when their subspaces overlap. In this paper, we propose a novel speech enhancement algorithm which uses deep neural network (DNN) to improve the encoding vector estimation of the NMF-based technique. A DNN is trained to represent the mapping from noisy speech to corresponding encoding vectors. The quality of the enhanced speech from the proposed NMF-based scheme adopting DNN-based encoding vector estimation is compared with that from the conventional NMF-based technique. The experimental results showed that the proposed speech enhancement algorithm outperformed the conventional NMF-based speech enhancement technique.

Full Paper

Bibliographic reference.  Kang, Tae Gyoon / Kwon, Kisoo / Shin, Jong Won / Kim, Nam Soo (2014): "NMF-based speech enhancement incorporating deep neural network", In INTERSPEECH-2014, 2843-2846.