Short answer: AI chatbots are bad at speaker design, and AI is very good at speaker design. Both of those are true. The difference is whether you're asking AI to remember an answer or to build something that calculates one.
Hours with a chatbot, no closer to a speaker
I'm seeing it more and more. Someone has a build in mind, they ask ChatGPT (or another chatbot) which drivers to use and what box to put them in, and a few hours later they're more confused than when they started. Then they email me. Just this week a customer told me they'd spent hours going back and forth with ChatGPT trying to pick drivers, and nearly went crazy.
It's not their fault. Here's what's going on.
Ask twice, get two answers
Try this. Open any AI chatbot, give it a woofer's model number and ask what box it needs. You'll get a confident reply with a volume, a tuning frequency and maybe a port size. Now open a fresh chat and ask the exact same question. There's a good chance the numbers move.
That isn't the chatbot being lazy. It's how they work. A chatbot doesn't look your driver up in a spec sheet and run the maths. It writes the most likely-sounding answer from everything it has read, and on speakers, what it has read disagrees with itself:
- Spec sheets change. Manufacturers revise their Thiele-Small figures between production runs, and old PDFs never die. Two "official" sheets for one driver can list a different Vas or Qts.
- Forums are full of half-remembered numbers. "I put mine in 60 litres and it sounded great" is a real data point about one person's room, not a box alignment.
- Similar model numbers get blended. The 8 ohm and 4 ohm versions of a driver, or last year's model and this year's, look nearly identical in text. A chatbot will happily mix them.
- Retailers copy each other's mistakes. One typo in a listing can show up on ten other sites.
Put all that in a blender and you get an answer that sounds right. Sometimes it is. Sometimes it's a box 30% too big with a port that chuffs at half volume. The problem is you can't tell which from the reply, because it reads exactly the same either way.
What AI is actually good at
Here's the bit people miss. The same AI that gives you a wobbly box volume is very good at writing software. And speaker box design is a solved maths problem. Give the Thiele-Small model a driver's measured parameters and a box, and it tells you the frequency response, cone excursion, port air speed and power handling. No opinion involved. Same inputs, same answer, every time.
That's what programs like WinISD have done for years. The difference now is that AI lets someone like me build that kind of tool around my own store, quickly, and check it properly.
So I did. I've built a modelling tool that does the same kind of maths as WinISD, with every driver I sell loaded straight from its spec table. It runs thousands of box simulations in about the time it takes to blink: every sensible volume, every tuning, every driver that might suit, and it throws out the ones that don't work. A port that's too fast (anything over about 17 m/s), a cone that runs out of excursion, a box that won't fit the space you've got. What's left is a short list you can actually compare.
The AI didn't make up any of those answers. It helped write the calculator. The calculator does the physics. That's the difference.
What a report looks like
Here's an example. It's the kind of question I get every week:
"I want a 12-inch sub for the lounge, for music and movies. I can fit a box of about 70 litres, ported. I'd like to spend up to about $600 on the driver. Which one, and what box?"
I ran it through the tool. Seven 12-inch subs that fit the budget, in 60, 70 and 80 litres, every port tuning from 16 to 40 Hz in half-hertz steps: 1,029 simulations, done before I'd finished reading the question. Then I went through the results and wrote up what I'd actually use.
The first page gives the answer straight up, then every other driver I looked at and why it didn't win. One of them, a great sub in its own right, didn't make the list at all: it simply doesn't work ported at this size. That's the tool doing its job.
Then the charts: how each sub plays in that box, side by side with my suggestion...
...and how loud each one can play before it runs out of cone travel, amp power or port. That second chart is the one a chatbot can't give you, and it's often the one that decides it.
Every report ends by saying exactly what the numbers do and don't cover. No hiding the fine print.
What this means for you
Right now: if you're choosing a sub, woofer, mid-woofer or mid-range, send me the driver (or the one you're thinking about) and what you want from it, and I'll model the box for you. It's free advice, the same as any question about parts. Use the contact page or the "Ask me a question" button on any product page.
I keep the full tool in-house and run it for you, because a box design is only as good as the questions behind it: what the speaker is for, the room, the space you've got, the amp. That's a conversation, not a form.
Coming in the next couple of weeks: the NB Speaker Builder on the site. It does the box modelling, similar to WinISD, and then goes further: it models the whole cabinet in 3D, bracing included, gives you the panel sizes, and gives you the option to have the panels cut for you. Pop your email on that page and I'll let you know the day it's live.
Tweeters are the exception to all of this. They don't have a box to design, so there's nothing for the model to do.
Where AI still can't help you: the crossover
This is the part I want to be really clear about, because it's where a lot of builds go wrong.
No AI, chatbot or tool can design a good crossover from spec sheets alone. Not mine, not anyone else's. Here's why:
- A driver measures differently once it's in your box. The spec sheet response is taken on a big flat baffle. Your cabinet's width, edges and driver position change it, sometimes by several dB right around where the crossover sits.
- Baffle step. As frequency drops, sound starts wrapping around the cabinet instead of going forward. On a typical bookshelf speaker that's up to a 6 dB drop in the lower mids, and the crossover usually has to account for it.
- Where the sound comes from matters. The tweeter and woofer sit at different depths, so their sound arrives at slightly different times. The crossover has to line them up, and that depends on your exact cabinet.
- Every driver is a little different. Two of the same model off the line won't match perfectly.
A crossover designed off the spec sheets is a guess. An educated guess, but a guess. To get it right you need to measure each driver in the finished cabinet: frequency response and impedance, with a measurement mic. It's cheaper and easier than it sounds. I've written a guide on measuring speakers on a budget with REW, and the basics of what a crossover does are in passive crossovers explained.
Once you've got real measurements, the maths takes over again, and that's where tools shine. A crossover designer is on the way for exactly that, and it'll only work from measurements, for the reasons above.
The short version
- Chatbots blend conflicting sources, so their speaker answers are inconsistent. Fine for "what's a passive radiator?", not for "what box does this woofer need?".
- Simulation tools (WinISD, and the one I use for customers) calculate the answer from the driver's actual parameters. Same answer every time.
- AI is great at building those tools. That's the right way to use it.
- Crossovers need measurements in the box. Nothing gets around that, AI included.
If you're stuck on any of it, send me a message. Choosing parts and talking boxes is the fun bit.