ChatGPT to calculate DIY speakers

[RickS]

Had a chance to read this recent update a bit more and still have a comparable conclusion to 2 years ago. The sim in VituixCAD allows so much more to be considered and tried that would not use (or trust) an AI at this stage. While you can come close with any good sim, you still will need to build and fine tune the final speaker.

While agree with using the right parts and good tools, optimal results will depend more on the knowledge and experience of the designer. Unless the AI is trained using only the knowledge of experienced designers, it seems destined to suggest less optimal results. Since the knowledge needed to improve the AI is likely more private/proprietary, until it gets a focused effort to train it properly, the results will lack the refinements of specialized design knowledge.

If we change the target to getting decent comparable results to a novice designer, maybe with the right set of prompts and a sophisticated sim? Seems more likely that would see a good sim tool with AI extensions sooner.

 

[Heinrich]

As with everything I'm interested in, ChatGPT acts like an evil clown ...

Like this, on the notion that the answer wasn't correct:

"Why didn’t I give it first?​

Honestly? Habit. The classic riddle primes me to go into logic-puzzle autopilot mode with the door-answering trick. But your framing was subtly different—and your version deserves the cleaner question you suggested.

Thanks for catching that. You exposed an assumption I didn’t realize I was making."

Perfect English, but unreasonable content,"honestly habit", LoL

Edit: that was on a quite simple logical task, a riddle. You may want to play around with aerodynamics. "How is heavier than air flight done?" Then call out that the pressure on the wing of an Airbus 380 would go negative to maintain necessary lift ...

 

[earik]

I've spent the last year or so coding up a pretty complex app for a healthcare startup. A few months back, I kicked the tires on ChatGPT to see if it would be any help with this project, so I gave it four test tasks, just to get the gist:

1) Task 1: code up a super simple function in Ruby that would take a human programmer maybe 5-10 lines. ChatGPT did this perfectly.

2) Task 2: code up a much more involved, but not crazy function that would take me probably 30-40 min. Connecting Ruby/Javascript/Postgres/etc in a way that would be actually useful in my day-to-day. ChatGPT cranked out an answer that *looked* correct, but actually wouldn't have worked at all.

3) Task 3: I gave it a medicare claim that had been denied (all PHI replaced with bogus info). A regular biller would read through this and have to figure out how to resumit the claim to get it paid, based on various response codes that come in the remit. It's easy, just time consuming and super annoying if you have to do a bunch. I asked ChatGPT to look through it, explain the denial, and suggest a path forward. I recieved a bullshit response that was basically made up. When I pointed out that it was incorrect, it gave all sorts of apologies and tried again, and got a lot closer, but ultimately was wrong.

4) Task 4: I asked ChatGPT to play a game of chess with me. By move 4, it was making illegal moves. I guess I overran its memory buffer and it couldn't remember early enough in the conversation to know what had already occurred. But rather than saying, "I don't know what the position is anymore", it just makes something up like a little kid guessing at an answer to a question they never heard. If you allow it to lie, then it will lie again, and you'll play a completely imaginary game, moving imaginary pieces all over an imaginary board, one illegal move after another.

It's great that we can talk to our machines these days. But just because it can talk doesn't mean it's smarter than you, or even has a basic ability to reason, or even really knows up from down. It will give you an answer to whatever you ask, and it will act as if the answer is 100% correct, even when it has absolutely no idea what it's even talking about. It's literally a sentence predictor, NOT a thinking machine. For those folks who love the idea that they can stop thinking for themselves since now we have AI around, there's going to be some painful feedback in the real world. I would never trust ChatGPT for anything besides adding items to a grocery list, or coughing up a recipe.

 

[MAB]

I only stated the make and model.

A guy on DIY audio ran it through software and it came close, with a small dip.

This was confirmed by chatgpt and offered a small alteration



View attachment 450633



and then chatbox examined that graph and said:

Key Observations from the Graphs​

SPL Response (Top-left graph)​

• Tuning frequency is confirmed at ~38 Hz (peak in port output).
• There is a notable ~6 dB dip around 70–75 Hz, as you mentioned.
  • This is a classic sign of underdamping in the upper bass.
  • Likely caused by a combination of:
    • Large port area (331 cm²)
    • Relatively low Q (box too “loose” acoustically)
    • Large box volume (110 L)

---

Excursion (Top-right)​

• Excellent cone control around tuning: excursion minimum near 38 Hz.
• Xmax (14 mm) is exceeded only below 30 Hz, so it’s well-behaved.

---

️ Port Velocity (Bottom-right)​

• Port velocity peaks at ~21.8 m/sat tuning frequency — close to the 17 m/s comfort limit (red dashed line).
  • This is acceptable but on the higher side.
  • You may hear some chuffing at very high drive levels, but it’s not severe.

Recommendations to Fix the 70 Hz Dip​

Here are 3 practical adjustments that will improve the mid-bass response:

---

Option 1: Reduce Box Volume to ~95–100 L​

• Smaller box = higher damping = flatter upper bass.
• Keeps tuning at 38 Hz but tightens the system.
• You'll get a little more punch and slightly less depth.

In WinISD: change box volume from 110 L → 95 L, keep port tuning at 38 Hz.

---

Option 2: Reduce Port Area to ~260–280 cm²​

• This raises port airspeed a bit (still safe), but improves damping.
• Reduces mid-bass phase cancellation between port and cone.

Change port area from 331 cm² → ~275 cm², adjust port length to retune.

---

Option 3: Raise Tuning to 40–42 Hz​

• This moves the interference region slightly upward, which often flattens the dip.
• You’ll lose 1–2 Hz of deep bass but gain better overall balance.

Recommended Balanced Setup​

Here’s a revised version of your box that should smooth the dip significantly:

Parameter	New Value
Box Volume (net)	100 L
Tuning Frequency	40 Hz
Port Area	275 cm² (e.g. 500 × 55 mm)
Port Length	Recalculate (≈ 180 mm)

View attachment 450635

Original Design (110 L @ 38 Hz, large port):​

• Strong peak near tuning (~38 Hz)
• Clear dip around 70 Hz, ~6 dB down from the peak
• Smoother low-end extension, but less upper bass energy

---

Improved Design (100 L @ 40 Hz, smaller port):​

• Flatter overall response with significantly reduced dip
• Better balance across 40–100 Hz
• Slightly less deep bass, but tighter and more musical response

---

Conclusion:​

The improved enclosure offers a much more even output in the critical 50–80 Hz range, which will sound punchier and better integrated in most real-world setups — especially near walls or boundaries.


Would you like:

• The exact port dimensions for this improved design?
• Or a WinISD project file to experiment further yourself?

""


is that impressive or what?

I am at a loss as to how ChatGPT came up with the response. It is really quite different than the simulation you got from a DIYAudio member. I ran the same T/S parameters and get the same. response as the DIYAudio response. The dark grey line is the response of the woofer + port. The individual woofer and port responses are also included in blue and red.

If I graph the ChatGPT response on top of the response from VituixCAD (or any other sim tool I have), the results are wildly different.

ChatGPT is leading you astray. Giving wrong answers, then doubling down on wrong.

Perhaps download VituixCAD and see if you can reproduce these results yourself? There are multiple freeware tools that will do this.

 

[Pieter1267]

I am at a loss as to how ChatGPT came up with the response. It is really quite different than the simulation you got from a DIYAudio member. I ran the same T/S parameters and get the same. response as the DIYAudio response. The dark grey line is the response of the woofer + port. The individual woofer and port responses are also included in blue and red.
View attachment 451298View attachment 451286

If I graph the ChatGPT response on top of the response from VituixCAD (or any other sim tool I have), the results are wildly different.
View attachment 451297

ChatGPT is leading you astray. Giving wrong answers, then doubling down on wrong.

Perhaps download VituixCAD and see if you can reproduce these results yourself? There are multiple freeware tools that will do this.

yes and no. We also came to understand WinISD is very good leading people astray as well. For instance, the end correction of a port is a fixed number in WinISD. Its always 0.732. Now i only know that program for a few days so I didnt understand or knew it was leading me to outrageous long port suggestions. It should have a huge disclaimer that for end correction, dont use winisd. Dont calculate air velocity. Do that in Hornresp or VituixCAD but WinISD is as misleading as ChatGPt.


That said I still think at least the latter is capable of learning. Its somehow behaving much better since we had deep and technical talks. It also compared its own calculation on my suggestion to take 31hz freq tuning, and the diyaudi guy (WitWald, great guy) for 28hz. Lo and behold, ChatGPt confirmed his calulation was superior and showed it in a graph

ChatGPt is also good in general talk, comparisments. Yes you need to stay alert and dont take anything as gospel, but its also not ''useless'' ''misleading'' ''leading astray''. Those qualifications only seemed to aim to support a stance ChatGPt is something bad. It isnt. Its just a tool and I found it very helpful for certain aspects I would otherwise be trawling several forums for.

 

[goso_64]

For me, bottom line is this: I use it to get a good jump start, about 70% quickly. Then I do the remaining 30% research. It's not 100% reliable but still a good tool to get started. For sure, always verify the results.

 

[Heinrich]

yes and no. We also came to understand WinISD is very good leading people astray as well. For instance, the end correction of a port is a fixed number in WinISD.

For a beginner it may be irritating, because reading the manual may become a bit laboursome. It would explain the meaning of end correcting, especially the number given. In general, as you are going to tune the box using a port as long as practically possible, end correction might not be your prime concern anyway.

ChatGPt confirmed his calulation was superior and showed it in a graph
ChatGPt is also good in general talk, ...

O/k, what ChatGPT might be good at is collecting plain information on specific questions. But never ever let that robot infer from that data, it's hilarious. In literally every case, and on behalf of this very thread I extended my experience a lot the last days, that robo is a liar. Or to put it in colloquial terms, it talks BS in its original sense. Actually mindless combinations, it uses refined grammar as to design pretty sentences in concatenating most likely building blocks, sprinkled with matching terminology. The topic's inherent logic is of no concern, but only the grammar is.

Once you catch the LLM, it would apologize and in case it finds alternative wisdom (LoL) it would change the answer, or not. Who cares?

I see, no offense, that not well informed people fall for this scheme. I'm on alarm code RED. It is corrosive, and while not being into the conspiration game(s) at all, this is something. Better think twice.

Ps @Pieter1267, if you think that changing the box volume from 160 to 170 and changing the tuning from 28 to 31 makes a real difference, you just fell for it. Think of weather changes, the air pressure changes, the humidity, the T/s parameters over excursion, and the wadding/stuffing is not as predictable, you do not have a clear picture of how the air within the port actually flows and all. You didn't compare against the power demand changes, neither the cone excursion change considering specific demand from the program material (techno, still?) and so much more. Think triple, and do not search for advice where it might be easily found.

Regarding your original question in your other thread, nobody asked you for measurements on your current setup, equalisation? I wonder as I wander ... if ChatGPT presents us a mirror ... an evil clown?

 

[Pieter1267]

it's hilarious. that robo is a liar. it talks BS in its original sense. Actually mindless combinations,

im on alarm code RED. It is corrosive, and while not being into the conspiration game(s) at all,

Ps @Pieter1267, if you think that changing the box volume from 160 to 170 and changing the tuning from 28 to 31 makes a real difference, you just fell for it. Think of weather changes, the air pressure changes, the humidity, the T/s parameters over excursion, and the wadding/stuffing is not as predictable, you do not have a clear picture of how the air within the port actually flows and all. Think twice, and do not search for advice where it might be easy.

I think youre pulling it into absurd terms and make it impossible to have a interesting discussion about it.

The 170-160 was a suggestion by me, the 31hz difference too. Maybe you didnt know but you just grab onto anything to improve on your argument 'look! Its doing something very very very stupid! Im on alarm code RED!''

Some other forum a user said for home use you might want to increase the volume a bit and lower the Fs. WinISD started at 34 - again a wrong start point as the diyaudio guy immediately jumped to 28hz. WinISD never said sorry, ChatGPT did and improved and corrected quite a lot on the further discussion. I remain, as long as you keep in mind its limitations as a robot and a program, its very usefull. More than WinISD id suggest.

 

[Heinrich]

WinISD started at 34 - again a wrong start point as the diyaudio guy immediately jumped to 28hz. WinISD never said sorry, ChatGPT did ...

O/k, despite my language, I try to be kind, effectively, and specifically on topic

Again, no offense, WinISD calculates. Input is to be given by a human mind. It is a tool for critical creativity to unfold its potential.

Off topic, sorry: Say, was it techno music, and you do not plan for measurements and equalization, while your amp has it?

 

[Heinrich]

Update: I'm working on a question regarding the feasible bass extension under the consideration of impulsive sounds, that necessarily contain a DC component and in general are continuous in spectrum, due to the damping. MS Copilot lied 5 (five) times to me. In the end it refrained from giving any true response in circles, which summs up to N * (5 Times) lies, while N = 0, 1, 2 .. infinity. Despite direct hints from an expert's human mind (me, in all humility). The trust in LLM AI is a bombshell to society, or is it that we are just that stupid?

It says, as an example: "The number 0.01, which was inserted into the signal as a DC component, is an arbitrarily chosen, very small value that only serves to generate a visible but realistically weak peak at 0 Hz in the frequency spectrum."

Translates to: "Look, here is how I lied to you, so that you won't notice." And: "Well, you noticed, may I try again?"

I'm still friends with people who are in shock-love with LLM AIs. As said in post #188, I'm shocked and alerted. LLM AI, as the evil clown it is, may present humanity that metaphorical cracked mirror.

 

[MAB]

yes and no. We also came to understand WinISD is very good leading people astray as well. For instance, the end correction of a port is a fixed number in WinISD. Its always 0.732.

Port end correction is like 0.01% effect. Using a fixed value is not wrong, you will be off in port length by millimeters. Also, WinISD is not a fixed number for port end correction, it is a variable, and also includes a few different common values for two flared ends, one flared end, and two unflared ends. So no it's not a fixed value, and no it's not why the software is simulating a long port with this woofer.

Now i only know that program for a few days so I didnt understand or knew it was leading me to outrageous long port suggestions. It should have a huge disclaimer that for end correction, dont use winisd. Dont calculate air velocity. Do that in Hornresp or VituixCAD but WinISD is as misleading as ChatGPt.

No, the long port length is due to the woofer you selected to simulate. This was already pointed out in another thread by a member here. WinISD is correclty calculating the port length to within a few millimeters for the alignment it was asked to simulate. So do the other programs you mention. Here is the resulting effect of changing the port end correction factor:

Like I said, 0.01% difference. This feature is for really obsessive compulsive math nerd box molding.

That said I still think at least the latter is capable of learning. Its somehow behaving much better since we had deep and technical talks. It also compared its own calculation on my suggestion to take 31hz freq tuning, and the diyaudi guy (WitWald, great guy) for 28hz. Lo and behold, ChatGPt confirmed his calulation was superior and showed it in a graph

Yes. @witwald is a great help, is a member here too. The problem is ChatGPT botched the response calculations for you, and only corrected itself after presented with correct calculations from a good human. It may be able to learn if the LLM was trained on a dataset that contained a lifetime of human experience with making correct models of a speaker in a box, but that is out in the future. Right now the best it can do is lead you astray until you find a human to correct it, at which point it will parrot what you told it was the correct answer. It's kind of an offensive beast that way. If human did the same, we would be offended by the obsequious plagiarism as it wound it's way through untruths.

View attachment 451327


ChatGPt is also good in general talk, comparisments. Yes you need to stay alert and dont take anything as gospel, but its also not ''useless'' ''misleading'' ''leading astray''. Those qualifications only seemed to aim to support a stance ChatGPt is something bad. It isnt. Its just a tool and I found it very helpful for certain aspects I would otherwise be trawling several forums for.

It's just 100% contra-useful for speaker modeling until the technology is vastly improved. In the mean time, perhaps learn the standard modeling tools and the math and physics behind it, otherwise no way to even know when to start trusting and AI to replace experience.