Spectral Refinement

Spectral refinement in its initial version is a postprocessing stage that follows a conventional frequency analysis (e.g. a Fast Fourier transform with a weighting sequence such as a Hann window). The principle idea of spectral refinement is to refine short-term spectra by means of computing additional frequency supporting points in between the orignal points using a linear combination of the current as well as a few preceding and successive short-term spectra.

Spectral refinement can easily be implemented using short FIR filters in each subband. This results in a very low computational complexity.

Application Examples

In speech processing spectral refinement has been applied successfully as a pre-processing stage for fundamental frequency estimation as well as for echo cancellation. Evaluations have shown that pitch frequency estimation methods can significantly be improved for various SNR conditions. Echo cancellation experiments confirm that spectral refinement can enhance the performance in terms of improved steady-state convergence.