[FFmpeg-devel] [PATCH] lavc/lpc: exploit even symmetry of window function

[Rostislav Pehlivanov]

On 10 March 2016 at 01:55, Ganesh Ajjanagadde <gajjanag at gmail.com> wrote:
>
>
> In fact, I personally have no idea why it is absolutely critical that
> a Hamming window is used, see 0cfdaf45c4 where Welch was used
> previously. At a first glance I think one can find a window that works
> and avoids libm, e.g Parzen or other polynomial windows.
> Rostislav, any thoughts?
>
>
Yes, though experiments the Hamming window was determined to work the best
with the TNS system. The Welch window was far too wide and caused
artifacts. The Prazen window seems too focused to let the LPC algorithm do
its job on correctly giving proper coefficients near the ends of the
window. The LPC coefficients are used by the decoder to shape the spectral
coefficients right before iDCT and after everything else, so it has big
implications on the quality.
I'll experiment with other types of windows, but I think the performance
penalty here is really worth it.