AMPLITUDE INDEPENDENT FEATURE EXTRACTION FOR EFFECTIVE SPEECH RETRIEVAL

The performance of speech retrieval systems is measured by its efficiency, accuracy and
noise robustness. In existing approaches, the acoustic features are extracted from the
power spectrum of the speech signals. These acoustic features are extracted from the
amplitudes of the power spectrum of a speech signal. It leads the system to be dependent
on the amplitudes for feature extraction. This amplitude dependency is one of the major
limitations, because these amplitudes can be easily varied by the quality of input device,
microphone position and in the presence of environmental noise, which degrade the
system performance. To overcome this problem, a novel approach of amplitude
independent feature extraction for noise robust speech retrieval is presented. Singular
values are used as acoustic features in this approach which are extracted using singular
value decomposition. These singular values are amplitude independent and noise robust
features. This work is motivated by the fact that, when a signal is SVD transforms its
singular values decrease abruptly with rank increment and data takes a form in which
first singular values contains the maximum amount of signal data