Refinement Methods as Preprocessing or Postprocessing for Short-term Spectra Analysis

In this contribution a generalized spectral refinement method of speech signals will be presented which can either be applied at a preprocessing stage before or at a postprocessing stage after a conventional frequency analysis. The refinement approach is based on a linear combination of weighted and shifted input speech signal segments. By applying the refinement method either in time or in the subband domain a refined version of the original short-term spectrum is obtained. Furthermore, the relationship between a polyphase filterbank and a spectral refinement technique will be considered. It will be shown, that the proposed refinement method applied as a preprocessor of conventional DFT-based analysis filterbanks is an extension of polyphase-based filterbank structures. The proposed spectral refinement technique is also very efficient and requires only few multiplications and additions. Experimental results based on the proposed refinement methods have shown that the robustness of pitch frequency estimation can be enhanced considerably.