The Application of Blind Source Separation to Feature Decorrelation and Normalizations

by Laura, Manuel

Abstract (Summary)
We apply a Blind Source Separation BSS algorithm to the decorrelation of Mel-warped cepstra. The observed cepstra are modeled as a convolutive mixture of independent source cepstra. The algorithm aims to minimize a cross-spectral correlation at different lags to reconstruct the source cepstra. Results show that using "independent" cepstra as features leads to a reduction in the WER. Finally, we present three different enhancements to the BSS algorithm. We also present some results of these deviations of the original algorithm.
Bibliographical Information:

Advisor:Patrick Loughlin; Luis F. Chaparro; Amro El-Jaroudi

School:University of Pittsburgh

School Location:USA - Pennsylvania

Source Type:Master's Thesis

Keywords:electrical engineering


Date of Publication:10/03/2006

© 2009 All Rights Reserved.