This has always puzzled me. Even though, over the past years - there have been a couple of papers published on the subject - none of them sound particularly great - tuning data is almost always missing. Why is reverb algorithms such well kept secrets?

Is it because there are very few companies that actually do reverb?

Is it because, in a sense(take with a pinch of salt) the only thing digital audio can do better by default than analog?

Is it just too much work?

It is the work of the Illuminati and their dark arts.

Oh and for anybody interisted in the subject - this is probably the best thread on the net:

http://www.gearslutz.com/board/geekslutz-forum/380233-reverb -subculture.html

Andrew

Try to read as much as you can of what Sean Costello (aka valhallasound) has written on the subject. Some links:

http://freeverb3.sourceforge.net/reverb.shtml
http://www.kvraudio.com/forum/viewtopic.php?p=3106441&highli ght=dattorro#3106441
http://www.kvraudio.com/forum/viewtopic.php?t=236590

Tuning data may or may not be missing, but if you are only using the ideas from a single paper and expect good results based only on that, that's going to be a disappointment in any case. You should at least be combining a good sounding early reflections model with a good sounding late/diffuse reflections model.

I would estimate the time used regarding the development of a high end reverb :

30% - The actual algorithm - the topology
70% - Tuning of the various values

I'll have to admit that I haven't found any papers that actual have a decent reverb or explain the various known high-end algorithms.

Every body can create a FDN reverb with prime numbers, but I have yet to hear any really good ones.

I missed this gem in my search, because it was in "effects" rather than the DSP section: MVerb - open source reverb

That's a classic in Reverb - and one of the few "complete" reverbs ever explained. Most people will know it as Dattorro's plate reverb, it's explained in Jon Dattorro's Effect Design - Part One . Basically the one he got sued over by Lexicon - Dattorro won btw

I'd say 40% topology, 70% tuning. The problem is if either one is off by a bit - the whole dang thing falls apart.



The best sound I could ever get out of an FDN was with a somewhat heavy modification:


 LP -> Allpass -> Delay. in a 16x FDN

Sum both the allpass(1/3time) and delay(2/3time), Modulate Allpasses. Used a sparse matrix.



Was much better, still not great, never pursued the idea further though.

Tuning data is ihmo missing for two reasons. First and foremost, it's not required (perhaps not even desirable) in research papers - which mostly deal with ideas, rather than actual implementations. Second the authors may not have the best data available, or they want to keep it for their own use later.

Richard

As far as I can tell by listening it's not complete. Trying to fine tune the parameters doesn't really improve things either, the default values on the paper are fine. It sounds better when combined with an early reflections model, however.

I did add a basic early reflection model into mverb if you want to take a look at the code. The early reflections aren't fantastic (just a basic multitap delay) but it does add a bit to the sound. I didn't spend much time tuning them so there is room for people to improve the code however

In spite of being simple that's still good enough to hide a few problems, resulting a better sounding reverb. My test tone was muted guitar strings strummed. That kind of percussive sounds reveal interesting problems that just don't get fixed by parameter adjustments, so the model needs to be backed by something that doesn't have the same problem.

Echo density has always been one of the main problems with most reverb. Though - taking the amount of resources it uses - the classic figure-eight is an excellent performer in my book. A cheap addon-trick (assuming different L/R times in the ER multi-tap) is to take about a third of the left side multi-tap -> pass it to an allpass -> and then sum with the opposite-side, ditto for the right. You could also just simply multi-tap delay and cross-sum - my Sony D7 has some small verb patches that uses the same trick. Personally I prefer the cross-diffusion trick more though. Also - always make sure that the Early Reflections and the Late Reverb join together smoothly.

Regards
Andrew

Is there anyone who has tried this? http://www.convexoptimization.com/wikimization/index.php/Dat torro_Convex_Optimization_of_a_Reverberator