I’m a bit late, but your discussion reminded me of this paper:
I. Kauppinen and K. Roth, “Audio signal extrapolation - theory and
applications,” Proc. of the 5th Int. Conference on Digital Audio
Effects (DAFx- 02), 2002.
The idea is to extrapolate in time-domain using autoregressive process model.
Each sample is modeled by
x(n)= ∑a_k * x(n−k)
with {a_1, a_2, ..., ap} being the AR coefficients. The AR coefficients can be
identified by system identification algorithms like the Burg algorithm.
Basically to extrapolate W samples based on ns past known samples:
• Identify the AR coefficients by using the Burg algorithm
• Initialize the filter with ns past known samples just before the section to
be extrapolated
• Feed zeros of length W into the filter
The quality of the extrapolation can be very good, even though this is
computationally quite expensive.
I’ve reimplemented it in python:
https://github.com/faroit/freezefx
Also have a look at this notebook for some audio examples:
http://nbviewer.jupyter.org/github/faroit/pyfreeze/blob/master/freeze_demo.ipynb
> On 16 Sep 2016, at 21:59, Emanuel Landeholm <emanuel.landeh...@gmail.com>
> wrote:
> Simple OLA will produce warbles. I recommend a phase vocoder
>
Compared to OLA or FFT based methods, the AR model extrapolation sounds quite
natural (especially when using higher filter orders).
Cheers
Fabian
_______________________________________________
dupswapdrop: music-dsp mailing list
music-dsp@music.columbia.edu
https://lists.columbia.edu/mailman/listinfo/music-dsp