posted 09 July 2001 15:13            

Along the lines of audio/signal processing, it occured to me that a useful addition to Kyma would be a histogrammer from which things like medians, means, stdevs, and so on could be obtained about signal statistics.

For noise removal systems, especially, this would be useful in helping to tease apart the contributions from noise and signal.

...one for the wish list...

Cheers!

posted 09 July 2001 18:02            

What I had in mind was a real-time histogrammer, akin to the FFT's in Kyma. They could operate on (overlapping) blocks like the FFT of some specified size.

That would eliminate the problem of continuous removal of the oldest sample when new ones are added. And there wouldn't need to be a huge buffer that holds samples, needing constant recomputation.

They might provide a (logarithmic) graphical display with live update -- that would allow us to see bimodal or multimodal distributions if they occur. But more importantly, I would like to be able to retrieve a mean, stdev, and median from the main mode of the histogram.

A median estimate would give a robust measure of the dominant noise level, for example. The stdev would be needed to help interpret a mean.

I suppose I can already compute a running mean and variance using integrating delay lines. But I can't get a median this way, or the 10 and 90 percentiles, etc.

Perhaps, if it simply spewed out the histogram bins (specifiable bin widths or overall width for fixed number of cells), then we could use a different kind of running Max that looks instead at the abscissa where that max occurs (MaxWhere?) instead of the ordinate of how large it is.

The graphical display could then be handled by an oscilloscope display.

I should give some more thought to this, but we use realtime histogrammers all the time in image processing. Mostly for establishing noise levels and bounds on the noise variations. We actually tend to use medians and MAD's more often than not, since these are robust against outlyers. (MAD = median absolute deviation from the median, the corresponding measure to a stdev)

(Hmmm... what would a histo-equalized sound be like? In images it stretches the contrast so you see more detail...)

[ Actually, if I had a spewing histogrammer and a MaxWhere, then I could compute the MAD using delayed inputs, subtracting the median, and computing another median from the absolute differences with another spewing histogrammer...

Is this getting bizarre enough? Or should I also mention the desire for 2-Dimensional FFT's? ]

- DM