As-is, as Dave & others have said, it's in effect electronically written junk, or more charitably a 1st pass 'guesswork' filter akin to what you'd have done from a magazine guideline up to -oh, call it about 2000, if you had no measurement equipment available and no access to any of the early crossover design simulators like CALSOD, Netcalc &c.
However good the 'AI' (I'm dubious on that name, but we're not likely to get shot of it now) software is, unless it incorporates dedicated design programmes, can use them, and is also provided the same data that anything, human or silicone requires for a high-quality simulation, i.e. full box dimensions, in-box impedance and on / off axis frequency & phase responses -it can't do any more, however much verbiage it outputs. It doesn't have a crystal ball.
However good the 'AI' (I'm dubious on that name, but we're not likely to get shot of it now) software is, unless it incorporates dedicated design programmes, can use them, and is also provided the same data that anything, human or silicone requires for a high-quality simulation, i.e. full box dimensions, in-box impedance and on / off axis frequency & phase responses -it can't do any more, however much verbiage it outputs. It doesn't have a crystal ball.
Lately I have been trying to use AI's to code. Generally they are like talking to someone that knows every version of a language but does not know what is relevant to the question. They often suggest functions that do not exist or existed in the past.
While AI's have saved me some time they have also cost me just as much.
I would not use one to design a speaker.
While AI's have saved me some time they have also cost me just as much.
I would not use one to design a speaker.
If you would like me to delete the thread, I am more than happy too. But the addition of reasoning in these new models is rather amazing (I posted without really digesting the output of the above info, was excited that is was reasoning with itself and adjusting parameters to 'work' better, questioning port length for example), and yes there output is still poor, but I do think the progression of this tech will change almost every facet of our lives, in the copy/paste it missed off all the formulas that made it look smart.
I do hear ya'll. I'm finishing up some boxes to house some Hypex Fusion amps so I can build up a decent active two way having used full range drivers for ever. Interestingly I did get the new o3 mini to guide me through the basic fundamentals of crossover design, happy to share if anyone wanted to have a review, I found it interesting to be able to ask it how a Butterworth had a -3db roll in, or why a certain part was used . I have a few books on speaker design, but tend to get lost when I see equations like this...plus I hate soldering, so active is a preferred root for me.
The new models are now able to write scripts etc to model fluid dynamics and I don't see it being long before they could excel at speaker design, it offered a prompt to add the modelled data into VituixCAD. I don't think it will be long before it could fully model in room responses and suggest improvements.
So, just for fun until this is able to replace all humans, what would it take for the models to impress?
@Scott. How do you go about designing a speaker (2 way monitor of instance) as a process I would love to see how close ai can come to being good at this process.
I do hear ya'll. I'm finishing up some boxes to house some Hypex Fusion amps so I can build up a decent active two way having used full range drivers for ever. Interestingly I did get the new o3 mini to guide me through the basic fundamentals of crossover design, happy to share if anyone wanted to have a review, I found it interesting to be able to ask it how a Butterworth had a -3db roll in, or why a certain part was used . I have a few books on speaker design, but tend to get lost when I see equations like this...plus I hate soldering, so active is a preferred root for me.
The new models are now able to write scripts etc to model fluid dynamics and I don't see it being long before they could excel at speaker design, it offered a prompt to add the modelled data into VituixCAD. I don't think it will be long before it could fully model in room responses and suggest improvements.
So, just for fun until this is able to replace all humans, what would it take for the models to impress?
@Scott. How do you go about designing a speaker (2 way monitor of instance) as a process I would love to see how close ai can come to being good at this process.
Last edited by a moderator:
Gentleman,
Can AI help to design a better loudspeaker in order to avoid 1000+ user iterations in VCad?
There is at least one adaptive tool freely available that goes way way beyond the classic optimizers to do so, so yes, it can and has been demonstrated on the very pages of DIYA. But maybe that tool should not be called AI...
What of course is needed as input is a set of files: the relevant .zma files of the drivers and SPL measurements -the infamous 19 (5 degree increment) measurements horizontal and vertical as described by Kimmo- of the drivers involved in .frd or .txt format..
Look and ye shall find....
Can AI help to design a better loudspeaker in order to avoid 1000+ user iterations in VCad?
There is at least one adaptive tool freely available that goes way way beyond the classic optimizers to do so, so yes, it can and has been demonstrated on the very pages of DIYA. But maybe that tool should not be called AI...
What of course is needed as input is a set of files: the relevant .zma files of the drivers and SPL measurements -the infamous 19 (5 degree increment) measurements horizontal and vertical as described by Kimmo- of the drivers involved in .frd or .txt format..
Look and ye shall find....
Quite -Xmachina can be remarkably effective providing the 'real' designer supplies it with the necessary data & (although they aren't called that) the design goals -which for a properly optimised speaker they should know and be working toward from the outset. At the end of the day, it's 'just' (ha!) an extremely formidable optimiser of topologies as well as component values.
That's a tricky one, since if I (or anybody else, preferably more capable than I am, since Solomon I ain't & don't pretend to be) wrote a high quality step-by-step process, it would be essentially book-length, if it also incorporated extensive reasoning behind each & every step according to whatever design goals are set in advance, & particularly if it doesn't simply follow standardised / generic formulas or alignments for enclosure & filter design. And secondly -while I don't think the 'AI' term is necessarily what's often thought, my opinion of it is the same as it's always been, and up-front: I don't like it, and do not, under any circumstances, wish to have any voluntary involvement with it whatsoever. True -as soon as something's up on the 'net, then like it or not, it's pretty much game ( even if not entirely 'fair') as it's likely to get trawled at some point. That doesn't mean I want to make contributions to it though, especially when I don't know just how much it may or may not corrupt what I said in the future. If others enjoy it -well & good. My choice, when possible, is to let it pass by.
That's a tricky one, since if I (or anybody else, preferably more capable than I am, since Solomon I ain't & don't pretend to be) wrote a high quality step-by-step process, it would be essentially book-length, if it also incorporated extensive reasoning behind each & every step according to whatever design goals are set in advance, & particularly if it doesn't simply follow standardised / generic formulas or alignments for enclosure & filter design. And secondly -while I don't think the 'AI' term is necessarily what's often thought, my opinion of it is the same as it's always been, and up-front: I don't like it, and do not, under any circumstances, wish to have any voluntary involvement with it whatsoever. True -as soon as something's up on the 'net, then like it or not, it's pretty much game ( even if not entirely 'fair') as it's likely to get trawled at some point. That doesn't mean I want to make contributions to it though, especially when I don't know just how much it may or may not corrupt what I said in the future. If others enjoy it -well & good. My choice, when possible, is to let it pass by.
I Appreciate the response Scott. I fear it will wipe out 50% of the workforce around the world and will be a very different place in the years to come. Robots building robots.
For clarity I didn't mean the entire process, but more, concept - drivers - box optimisation - then XO, or Perhaps the brief requires the box to come first, then drivers then XO. I assume the XO always comes after you know the driver parameters, do you use the box design to alter response after XO?
@Boden I mean I'm still impressed you can get the models to give you the data specs in a format you want. And being able to speak in English to do commands and compilation that would take ages is impressive. We could get it to scan the net for all drivers and extract all the information, fact check itself and present it as a zipped up file of correct T&S parameters in a way all software could access...
For clarity I didn't mean the entire process, but more, concept - drivers - box optimisation - then XO, or Perhaps the brief requires the box to come first, then drivers then XO. I assume the XO always comes after you know the driver parameters, do you use the box design to alter response after XO?
@Boden I mean I'm still impressed you can get the models to give you the data specs in a format you want. And being able to speak in English to do commands and compilation that would take ages is impressive. We could get it to scan the net for all drivers and extract all the information, fact check itself and present it as a zipped up file of correct T&S parameters in a way all software could access...
Even the best AI can't design a proper x-over without knowing the impedance curves of the drivers (in the box!) and the SPL curves of the drivers (in the box / on the baffle!).
Box optimization is rather straightforward (and far less relevant than the diy community seems to assume..). X/O optimization is a very different ballgame, given the large number of available options i.e. design targets.
Here is a link to the chat. These are not just LLM's anymore and are thinking. Whether they are right is another matter. You can see the process which I find interesting.
https://chatgpt.com/share/67a1f24d-4ed4-8008-888a-3a797b15aaa9
I'lll see if I can run the script it has produced as it can't yet render the output as a graph.
https://chatgpt.com/share/67a1f24d-4ed4-8008-888a-3a797b15aaa9
I'lll see if I can run the script it has produced as it can't yet render the output as a graph.
I wouldn't be entirely surprised, put it that way. I'm not seeing this as a positive step forward either. 😉
I figured as much; to be honest the same answer applies to do a proper job of it.
Working on the assumption that this is not going to be fed into some AI software, commercial & DIY design are different fields & the process is usually going to have some major differences; from a DIY perspective (and then assuming it's for my own use), I'd normally have an initial target set of usage (including budget) & sonic goals set down first. From there, if it's not already known, I'd make a shortlist of drive units that will suit the requirements, & narrow the field down to those that will have the fewest compromises for the intended application. You can then do some more detailed work on the box alignment, proportions & construction, accounting as far as you can for probable effects e.g. diffraction etc. (sometimes doing a rough simulation in advance to reduce the potential for surprises), build the prototypes, take some basic measurements & if they're behaving as expected, move on to more involved; if not: investigate why not, & adjust as needed before going to the detailed measurements. From there, it's filter design through preferred method according to the objects intended, and again, adjust as needed. That's [very] roughly how I do it. I certainly wouldn't design an enclosure after the filter; for me anyway, details of filter design & adjustments almost always come last, with the minor exception of critical damping of the alignment when the speakers are in place, or any other changes have been made to the environment that affects it. But that tends to be fairly specific rather than a major effect on the whole spectrum.
Last edited:
It's not about intelligence power, it's about the procedure and data available AI needs to work on.
Human uses the actual impedance and FR curves taken from the actual box already built in order to design the crossover.
This is our basic procedure to consider electrical, geometrical and acoustic frequency/phase final response.
It's a super complex interaction of speakers and box made simple by simply measuring the final result with a mic with timed reference.
This information is not available before building the actual box, so AI cannot do anything here cause there is no data before building the speaker.
In order to just select speakers, define a box and position the speakers in theory (not actual building), we have to come up with a model for each speaker that can interact with a model for the box we intend to build (size, speaker positions, materials etc) and this model must simulate the final result including the electrical, geometrical and acoustic response (freq/phase)
If we have this complete model, then, we or AI could simulate the final result (crossover+speakers+box).
As far as I know, this model is not current available from any speaker vendor in the Internet, so data is not available for AI to work properly.
Human uses the actual impedance and FR curves taken from the actual box already built in order to design the crossover.
This is our basic procedure to consider electrical, geometrical and acoustic frequency/phase final response.
It's a super complex interaction of speakers and box made simple by simply measuring the final result with a mic with timed reference.
This information is not available before building the actual box, so AI cannot do anything here cause there is no data before building the speaker.
In order to just select speakers, define a box and position the speakers in theory (not actual building), we have to come up with a model for each speaker that can interact with a model for the box we intend to build (size, speaker positions, materials etc) and this model must simulate the final result including the electrical, geometrical and acoustic response (freq/phase)
If we have this complete model, then, we or AI could simulate the final result (crossover+speakers+box).
As far as I know, this model is not current available from any speaker vendor in the Internet, so data is not available for AI to work properly.
I was not talking about "optimizing the box", whatever that would be.
I was talking about the impedance curves and the SPL curves of the drivers in the box / on the baffle. These are pieces of what an xover needs to combine into something that sounds nice. If you don't tell the AI or whatever tool you're using, it won't know what it's working with.
Looking at the ChatGPT session posted above I think it is mostly a different way of presenting the information found online anyway. The formulas presented are generic and any tutorial written by a human would probably do at least as good a job of explaining the calculations and how they are done. Some people may prefer the "interactive" presentation style of chatting with an agent over reading a tutorial, but that doesn't mean that the answers are as good as you would get from an experienced human designer.
@mbrennwa do you mean real measurements or simulation’s.
Do we not think that o3 could operate like WinISD but with natural language?
@Scottmoose 🤖
Do we not think that o3 could operate like WinISD but with natural language?
@Scottmoose 🤖
Oh, joy. I've just looked at this and -frankly, I'm not sure why much value would be given since it's -well, not much, really. This is not 'advanced reasoning' -that seems a euphemistic term by whoever coined it when applied to a complex, multi-layered engineering & psychological situation that does not really have specific answers, and current generations of software have no ability to solve.
Blow by blow, and specifically reqested not for entering into 'AI' systems:
'It may be easier in practice to start with... then decide...' It may. But that is not the way you design an LR2 filter, which is a specific target / alignment, and on a flat baffle with direct radiating units will require additional measures to achieve a correct phase alignment & FR as a result.
'Slight tweaks may be required' aka 'now go and design it properly'.
Sorry, but this thing's value is roughly in proportion to a walrus on a summit ascent of K2. 😉 In fact, here's a walrus, on the basis that it's best to keep positive, and most things are improved by them.
Blow by blow, and specifically reqested not for entering into 'AI' systems:
It really isn't.
Or even 'acceptable' in many cases. Note the linguistic fudge already present.
'Proven design practices' according to whom?
In summary, the first is a linguistic fudge since I'm not sure who would call a crossover 'well-matched', but quality design depends on what you're trying to achieve. 'First-order for simplicity...' is drivel, since a decent 1st order can often require more, not fewer components, especially if no consideration is given to time alignment [natch]. The same also applies to LR2: especially if symmetric, if no consideration is given to time alignment & offsets, then performance will be lousy. And that's before you get to the heffalump in the room which is actual on-baffle frequency, phase & impedance responses, so all generic electrical filters are almost certain to go out of the window anyway. 'Ported' boxes do not 'boost efficiency' except in their own BW either. Since 'bookshelf size' is mentioned, it's interesting zero reference seems to be made to potential [not insurmountable, but exist] issues of a 'ported box' on a bookshelf.
- A 30mm dome is rarely described as 'small'.
- 'Detail' means nothing.
60's [sic]? It gets a heck of a lot lower than that. It contradicts itself further on. QED.
I do like a nice, clear 'blueprint'. 😉
- It doesn't help 'minimise diffraction'. It can, with genuinely careful location, not these vague statements, help reduce some of the effects.
- An acoustic waveguide is not a 'small vertical offset'. A waveguide is a waveguide, not a vertical offset.
- Waveguides are not typically used for the sole purpose of time alignment. They can be, but that's not normally their primary function, it's a secondary (often useful) attribute -depending on the listening axis, the offsets in the x, y and z axis, and providing the baffle is flat, because if the latter isn't, you've a whole new set of factors coming into play. Ye gods, why am I wasting my time on this dribbling twaddle?
More sillicone hand-waving. 'Hand matching', aka 'pair-matching' is certainly useful taken in general, but it's not normally done for the specific & sole reason of ensuring the 'overlapping responses sum flat with good phase coherence'. Since phase follows the FR in the piston and linear TL operating BWs (drive units being normally taken as minimum phase devices across that if nothing else), why this is elevated to some primary reason is mysterious.
- I've no idea what 'natural phase behaviour' or 'open' sound are supposed to mean. And since the MW19P in either impedance certainly does not have a particularly wide / smooth operating bandwidth, you are not going to get away with chucking a simple 6dB/octave slope at it.
- Why 'off-axis' response is taken as something different to 'free air' is another of life's mysteries.
Which is about as daft as you could get with an MW19P. You may indeed wish to adjust. Extensively. Or preferably, pick a midbass driver that's actually suited to this use -assuming you can find one. And account for acoustical offsets in all three planes while you're at it... 😉
The case rests m'lud. The 'standard formula' assuming static impedance, no account for actual frequency response or anything else for that matter.
Not bad advice, but non-polarised aside (and you can back-to-back a couple of polarised if you felt the need), not especially critical.
Ditto. As for the component, an air core isn't necessary. I'm a fan, but gauge is critical in the LF or the DCR will affect box alignment in particular by artificially altering the driver's effective electrical damping.
Worthless generic junk taking no account for actual requirements & simply fudging with a 'recalculate using your own...'. Note that it's 'designing' nothing so far, just giving a few bits of fairly misleading data with no reference to the actual situation or requirements that have been trawled off the 'net.
'If well implemented'. And this 'ain't the way to do that. Generic statement about LR2.
'It may be easier in practice to start with... then decide...' It may. But that is not the way you design an LR2 filter, which is a specific target / alignment, and on a flat baffle with direct radiating units will require additional measures to achieve a correct phase alignment & FR as a result.
- Curious this is coming after filter design (!)
- A ported box does not provide 'higher efficiency' except potentially in its operating BW.
- Not all birch plywood is created equal; like all generics, some is fairly poor.
Adequate, depending on tuning. 8ohm version marginal but just about acceptable in practice. Dodgy dimensions listed.
- 'Potentially reduce', not 'minimise.
- Yes. It should match. Or it won't fit. No reference to flush mounting.
- Upper section 'in front'? Que? Cutout should match; again no reference to flush mounting. Both a bit ironic when associated with prior references to rounding baffle edges to reduce diffraction'. It also doesn't know the driver dimensions, so again, hazy generic 'advice' rather than 'design'. 'A slight vertical offset' in 99.99% of cases does not provide 'time alignment'. You need something rather more dramatic than that, either in the physical or electrical domain.
Not in my book it doesn't, especially once series R from speaker wire loop, connection & series component DCR is factored in (there follows a bunch of standard equations trawled off the 'net).
- No reference to brace type or location.
- No reference to thickness, variations, or how to adjust.
Measure the internal volume? You should know in advance (!)
'Slight tweaks may be required' aka 'now go and design it properly'.
Yes, although WinISD can't help with crossovers. Measuring is also a good idea. 😉
Translation: 'no'.
Sorry, but this thing's value is roughly in proportion to a walrus on a summit ascent of K2. 😉 In fact, here's a walrus, on the basis that it's best to keep positive, and most things are improved by them.
Attachments
Last edited: