INTERSPEECH 2012
13th Annual Conference of the International Speech Communication Association

Portland, OR, USA
September 9-13, 2012

Effect of Speech Priors in Single-channel Speech-music Separation for ASR

Cemil Demir (1,3), A. Taylan Cemgil (2), Murat Saraçlar (3)

(1) TÜBITAK-BILGEM, Kocaeli, Turkey
(2) Computer Engineering Department, Boğaziçi University, Istanbul,Turkey
(3) Electrical and Electronics Engineering Department, Boğaziçi University, Istanbul,Turkey

In this study, we extend the catalog-based single-channel speech-music separation method such that it incorporate prior speech information to enhance the separation performance of the method. We developed the inference method that enable us to use the speech prior model which is obtained using pre-obtained speech signals. Complex Gaussian observation model which uses the Inverse-Gamma distribution as a prior model are used to develop the inference method. We compare the separation performance of the catalog-based method with and without prior speech model in both complex Gaussian and Poisson observation models. It is shown that for both observation models incorporating prior speech information improves the separation performance of the catalog-based method.

Index Terms: speech-music separation, prior speech model, speech recognition

Full Paper

Bibliographic reference.  Demir, Cemil / Cemgil, A. Taylan / Saraçlar, Murat (2012): "Effect of speech priors in single-channel speech-music separation for ASR", In INTERSPEECH-2012, 1235-1238.