Windows search

Summary
Signal weighting windows were examined for usefulness of results of spectral analysis for fundamental 
frequency (F0) measurements and for speech sounds recognition. 
There were applied two criterion: the power cepstrum value at the T0 point (the F0 period) and the 
variance of of an instantaneous spectrum as a dynamics measure of the spectrum. The dynamics was 
computed with reference to the spectrum after blind deconvolution and liftration. There were 
examined nine windows types, i.e.: Hamming, Hann, Keiser-Bessel, Blackman-Harris, Blackman-Nutall, 
flat-top, Gauss – two types, and rectangular. We examined windows of five widths each: 
64, 128, 256, 512 and 1024 points for a speech signal sampled at 16 kHz. 
The experimental material consisted of realizations of 6 Polish vowels. There were taken for this 
purpose 20 realizations of each vowel from every speaker (45) from Corpora data basis. Thus there 
were analysed as much as 6×20×45=5400 vowels realizations spoken by male (29), female (10) and 
children (6) voices. There was chosen the best part from each vowel realization for the analysis. 
As a criterion of these choices was considered value of the power cepstrum at the T0 point found 
using the Hamming window.
It turned out, that the best were Gauss windows with optimized standard deviation. The next, worse, 
was the flat-top windows for the dynamics emphasize and the Hann windows for the T0 measurements. 
The worst turned out to be the rectangular window. The best compromise – a window suitable for both 
tasks was the Blackman-Harris' window.
These conclusions were drawn on the basis of the ANOVA of repeated measures tests.

P.S.
We are really sorry that we could not attach all source data due to its huge size, but we will send 
it on request (quite for free).

We are still working with the problem. Next results will be soon.