Generalized refinement of short-term fourier spectra in time- and frequency-domain and its combination with polyphase filterbanks

In this contribution a generalized solution for the spectral refinement (SR) of signals is presented. SR essentially consists of a linear combination of weighted and shifted input signal segments (either in the time or the subband domain), yielding a refined (higher resolution) version of the underlying short-term spectrum. The proposed method is particularly suitable for speech processing systems with an existing analysis filterbank based on polyphase or DFT realizations. We first show that DFT-modulated filterbanks that incorporate SR in the time-domain can be seen as a specific realization of polyphase-based analysis filterbanks. Next, the polyphase-based analysis filterbanks will be combined through the proposed generalized form of SR, in the frequency domain. The resulting structure can be applied to refine either the frequencies of the short-term spectrum (or a subset thereof) or to refine the whole frequency range, including additional frequency supporting points. Lastly, as a potential application, we demonstrate the benefit of combining an existing polyphase filterbank with SR for the problem of pitch extraction/fundamental frequency estimation. Especially for lower fundamental frequencies, such combination can give a significant performance improvement compared to estimates based on the original signal spectrum.

ISBN:9781728153414

Jaar van publicatie:2019

Toegankelijkheid:Closed