ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Noise quality improvement through SVD equalization

Stylianos Bakamidis, George Carayannis

In the modern computer environments, special software is available for noise generation. Depending on the quality of the software, usually also related to the hardware, the spectral flatness of the obtained noise is generally acceptable for long sequences. For shorter sequences the open loop procedure usually utilized by the previously mentioned software fails to give the expected results. For number of samples less than one thousand the spectral flatness which is equivalent to the eigen value spread needs to be improved. A new method is proposed here based on the SVD analysis technique which leads to the improvement of the spectral flatness. For a given noise frame an algorithm is provided giving the possibility to set the eigenvalues to equal magnitudes which has as consequence the regeneration of a noise with increased spectral flatness. Simulation results gave significant improvement for both uniform and gaussian noise distributions.