Amos Storkey

Bayesian Fourier Transform Demos

At the moment I have one demo available. Suppose streaming audio was available, but it suffered from serious packet loss. Then we would want to be able to estimate as best we could what the missing data was from the data we had available. Suppose for example we were being sent this section of Bach Magnifcat in E flat major. Unfortunately due to an exceedingly bad network connection, what we actually received was a rather corrupted Bach Magnifcat in E flat major. Using the methods described on the Bayesian Fourier Transform page, with 2048 byte sections for each FFT, and using an empirical prior for each section based on a weighted decay of the posterior power spectra over the previous sections we obtain the restored data. It should be mentioned that this is an unoptimised demo: the data was not manipulated in any way prior to use, and the decay weighting parameters were set to a reasonable value. The choice of sections to be removed was done at random, ensuring that some sections of significant length went missing - this was the first setting and attempt I tried. Honest!


 
 
 
Amos Storkey 2000-2005.