7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Speech Enhancement Based on Generalized Singular Value Decomposition Approach

Gwo-hwa Ju, Lin-shan Lee

National Taiwan University, Taiwan

The singular value decomposition (SVD)-based signal-subspace approach for noise reduction has received high interests in recent years. With this approach, we can diagonalize the matrices constructed from noisy speech frames and divide the whole feature-space into signal-subspace and noise-subspace by the singular values obtained from the matrices. We then reconstruct speech from the signalsubspace only. In this way, speech signals can be successfully enhanced. This approach is very effective when the additive noise is white. If the noise is not white, we have to first whiten the noise spectrum prior to SVD-based approach and perform the inverse whitening procedure afterwards. Not only the process is complicate, but extra distortion may be introduced in such a process. In this paper, a generalized SVD (GSVD)-based approach for speech enhancement is proposed, which is useful regardless of whether the added noise is white or not. Experimental results show that this new approach can provide very good performance, specially better than the conventional spectral subtraction algorithm and SVD-based approach, in particular when the additive noise is non-white.


Full Paper

Bibliographic reference.  Ju, Gwo-hwa / Lee, Lin-shan (2002): "Speech enhancement based on generalized singular value decomposition approach", In ICSLP-2002, 1801-1804.