On Wed, Sep 29, 2010 at 11:33:37PM +0200, Philipp Überbacher wrote: In theory they'd go to minus infinity. What is worse is that they are also quite wide at the -3dB points. Just inverting it will produce a sum response that is more or less flat. It will have bumps of a few dB instead of the deep dips, but this can be corrected to some degree by adjusting the filter params. Ciao, -- FA There are three of them, and Alleline.

On Thu, Sep 30, 2010 at 05:40:46PM +0200, torbenh wrote: It would, but it would probably be overkill. We are not driving a set of speakers here, but three or four compressors, and there are good reasons to make the crossover quite gentle. That is my gripe with the FFT-based xover: it is far too steep. It could result in one note of a melody being processed by on compressor and the next, just a few semitones higher or lower, by another. This is probably not what you want, certainly not if those compressors are behaving quite differently which is the whole point of multi-band processing. I'd go for a filter type that is close to first order in the xover region and becomes more steep an octave or so higher or lower. This would produce some minimal (< 1 dB) ripple in the FR, but who cares if you are going to multiband compress anyway ? CPU usage would increase actually, if the filter is implemented really as a filter. For the linear phase thing, see below. 16k taps (at 48 kHz) will allow you to make most wanted types of audio filter (except steep highpass at low F), and zillions of very bad ones... I'd never advocate the use of such filters except if they are derived e.g. from an inversion procedure, as for room correction. Some of it. But it's being taken way over the top. There are (in audio) almost no good reasons to make a filter linear-phase, and many filters will sound very bad in their linear-phase form while being perfectly acceptable in their minimum-phase form. It's often claimed but not true that a linear-phase filter will 'ring' less. It will ring just as much as any other filter with the same amplitude response, and what is worse, half of the ringing will precede the main peak of the IR, which makes it much more audible. Let me give an simple example. You have a FIR of 4801 samples, all of them are zero except: Linear-phase form: F[0] = 0.02, F[2400] = 1, F[4800] = 0.02 Min-phase form: F[0] = 0, F[2400] = 1, F[4800] = 0.04 Max-phase form: F[0] = 0.04, F[2400] = 1, F[4800] = 0 All three have an amplitude response that oscillates around 0dB by about 0.35 dB. It is *very unlikely* that anyone would actually hear the effect of the minimum-phase form. The two others will have a 50 ms pre-echo of resp. -34 and -28 dB, which is easy to hear on some types of signal (anything percussive). Our hearing is very sensitive to anything preceding the onset or main peak of a sound: this is almost never a property of the original sound source but it indicates very often the presence of more than one of them. Which is essential info for any life form trying to avoid its predators, so this sensitivity may have its origins in evolution. Also all percussive sounds are 'minimum-phase' - they sound as 'ping' and not as 'gnip' or 'gnipping'. In cases where you need a long filter (that is for a FR that has much fine detail), the minimum-phase form will affect the envelope of a sound the least, and linear-phase is to be avoided. For 'smooth' filters it's OK, but I doubt very much if it actually improves the sound at all. Ciao, -- FA There are three of them, and Alleline.

On Sun, Oct 03, 2010 at 12:31:59PM +0200, Philipp Überbacher wrote: They are second order, using biquads. With first order this problem wouldn't exist as 1/(1+s) + s(1+s) = 1. What you see is not the shape of each filter, but how they combine. The lowpass is of the form 1/(1 + d s + s^2), the higphass is s^2/(1 + d s + s^2), s = j f / fxover, d > 0. The nominator of the sum doesn't have any term in s, and the two others cancel at the xover frequency where s = j. Hence gain is zero at the xover frequency. Just a small gain difference in the compressors. If you bypass them the dip is at exactly the set frequency. Ciao, -- FA There are three of them, and Alleline.

Hi Jan, For various reasons. One condition would be that the original authors agree that there is something to be fixed. This discussion started not about the IIR crossover but about the FFT-based filter. I presented the technical arguments, and the only response was a lot of obfuscation (by referring to my remarks as being about optimisation while I had made it quite clear from the start that this was not the case), irrelevant references to general theorems posted by someone who clearly does not understand them, and finally that 'three qualitatively more experienced DSP engineers' who remained anonymous and did not present any technical arguments were of the opinion that I was wrong. If there's something wrong in my analysis I'll be the first to admit it if presented with the evidence. If OTOH the response is to publicly attack my competence as a DSP engineer, then the one who does that should better have his own shop in order and know what he's talking about. Which as the IIR crossover example shows is clearly not the case (it's not a typo). Because before I post a single word of criticism I will have measured the application, analysed its code, and if necessary refreshed and updated my understanding of the relevant theory. A second reason is that there are multiple issues and fixing all of them would not be a matter of a few patches. And regarding the FFT-based filter, even if implemented correctly this remains IMHO the wrong kind of filter. It has its uses, but certainly not as a way to emulate a parametric EQ or a 30-band graphical EQ. If I would fix it, that would also imply I approve the way it is being used, and I don't. So, for the moment, 'I rest my case', and maybe I'll look into Jamin again in a few months. I did not start this discussion anyway. Ciao, -- FA There are three of them, and Alleline.

On Wed, Sep 29, 2010 at 12:19:09PM -0500, Jan Depner wrote: If you fix the filters then the IIR crossover would actually be the better one. The FFT-based filters are much too steep (this can be fixed as well of course). Ciao, -- FA There are three of them, and Alleline.

On Wed, Sep 29, 2010 at 8:43 PM, Patrick Shirkey &lt;pshirkey(a)boosthardware.com&gt; wrote: For the record: I don't know much about DSP. Steve did all the plugins. I just wrote the JACK I/O interface. --  joq

On Wed, Sep 29, 2010 at 06:43:42PM -0700, Patrick Shirkey wrote: Where did I do that ? For your reference, this is the original message in which I referred to Jamin's FFT-based EQ as a vocoder: <http://lalists.stanford.edu/lau/2010/05/0321.html> And when I wrote that remark - 'a vocoder being used as a filter', which is technically accurate - I just assumed that the developers would be aware of that, that it was their intention to make the EQ work in that way. They were certainly aware that the filter had it's problems and was not really working like a linear filter - that is why the overlap factor was increased from the original 2 to the much higher value (8 or 16) used today. Regarding the current discussion, it started with others quoting this message quite inaccurately, and me trying to correct these remarks. Followed by a (I think) a clear expose of what IMHO is wrong with this filter and why I call it a vocoder. Ciao, -- FA There are three of them, and Alleline.

On Thu, September 30, 2010 2:22 am, fons(a)kokkinizita.net wrote: Here's the greatest hits: "It's a good example of textbook DSP being applied blindly and without understanding the consequences. " "But it's a kludge, and a solution to a problem that shouldn't exist in the first place." "As said, there are two issues with the EQ. The first is that the implementation is subobtimal and introduces artefacts that have to be (and are to some extent) hidden by changes added after the initial release. The second is that this type of filtering allows and invites to do things that do indeed 'destroy' the sound from a purist POV, while adding nothing useful from any other POV." and I suppose this was just plain vindictiveness "Don't know where you get this, but I assume it's not the result of your own intellectual efforts and so I can just say 'bullshit' (the pure liquid variety actually). I won't even comment on the rest, as I'd have to be at least as rude :-) " Don't know why I got the impression you were being rude or anything... Must be just me. It is great that you have taken the time to explain the issues that you have spotted. As you rightly said I do not have the experience that you have and I appreciate you taking the time to explain in detail so that we could all have the opportunity to learn something new about jamin and DSP. -- Patrick Shirkey Boost Hardware Ltd.

On 09/30/10 15:57, Patrick Shirkey wrote: Some of these are quite amusing and IMHO most of them justified in the original context. ..but more importantly none of them was a personal insult. Which can not be said about some of your comments. If you want my take on it: pipe it into a text-to-speech synth and make a RAP song out of it: "Parental Advisory Explicit Lyrics." Bonus points if you manage to work Bob M's Jammin' into it and master it with jamin Come to think of it we might even start a petition for the Unplugged version of it featuring Fons at the next Linux Sound Night :) peace, robin

Philipp Überbacher a écrit : +1, I agree "100% X (y/w+(x2-AB3/4))" with the Bob Marley's cover masterd with Jamin... the name of this band is talking heads, OUPS ! The name of this list is LINUX AUDIO USER I meant... These days it looks like a flame trolling stuff; sure we get many knowledge shared in it (maybe not? as a matter of fact, don't care about maths, what I like is sound-and Jamin-)) but maybe some people will get more karmic bonusses with music practice instead of post here. Just my advice, with big respect for Linux gurus concerned (if any). PEACE. PS : Maybe, there are some lists around here, reserved to dev people, where.... SMILE, MEN !