Using AI to help with room correction

Rooms and Treatments
1

Yeah, I know, you can’t even discuss your favorite Mexican restaurant without someone mentioning AI. But it’s creating some opportunities very difficult to pass up. One thing I’m finding that it does well for audiophiles is consolidating and synthesizing information that is so far-flung in so many places, nobody could ever collect it all, no matter how much time you have. And most of what you need is FREE, which sure isn’t typical for audio improvements!

It all started when I read a Paul’s Post about upsampling to take advantage of filters in a DAC. I asked Google’s Gemini AI (the system you get when you click the “AI Mode” button in the search bar of Google’s search engine) whether the DAC I have - the PS Audio Stellar Gold - would benefit from upsampling. It basically said no, except maybe if I was using room-correction DSP, which I am. It then told me how to properly use Roon’s upsampling.

I then told AI that my main objective is to emulate the sound environment typically used when mixing or mastering a recording. I said that my journey had been to use Focus Fidelity’s evaluation and room-correction system to generate convolution filters similar to what are used for mixing and mastering rooms. For some reason I can’t recall, I had settled on Focus Fidelity’s Harman filter with a 4dB bass boost. Gemini didn’t like the filtering choice for my desires and gave incredible details as to why. The thing that AI can do so beautifully is take a problem with a lot of different variables and process it against virtually everything out there. So, now fully in the. rabbit hole, I spent about 2 hours writing up a detailed description of my room and all that went into building it, my system, what kinds of music I play most. It took my information and processed it against everything it could find about my details (in about 5 seconds) and said this:

"First, let’s acknowledge the incredible engineering achievement of your listening space. A double-wall “room-within-a-room” on a concrete slab, isolated sub-panels, decoupled turntable/NAS placement, and extensive geometric/acoustic treatment puts your physical room’s noise floor (28 dB) and reflection management ahead of many commercial project studios.

Your Vandersteen 2ci speakers—renowned for being time-and-phase correct—paired with the classic tube mid-range of your Convergent Audio Technology (CAT) preamp and the effortless headroom of the PS Audio Stellar M1200 monoblocks, create the perfect foundation for a mastering-grade presentation.

Because your physical room is already exceptionally well-treated, your Focus Fidelity Finetic configuration should be highly conservative. You are using DSP to apply a final “polish” rather than fixing catastrophic room issues."

It then instructed me on setting up the filters, analyzing the curve results - all based on my specific information. I even asked it if I could upload a screen shot of the Focus Fidelity results graph so it could look at it - IT DID. Then it told me how to make sure the filters were correctly installed into Roon, adding a headroom filter just in case.

Were we finished? Heck no. It asked me what music I would be using to evaluate the change, and suggested some tracks of its own. The frosting on the cake was when it suggested a particular recording of Copland’s “Fanfare for the Common Man,” I told it that when I was in a youth symphony in high school, we called it “Fanfare for the Common Cold.” And for the rest of the session, it referred to it as “Fanfare for the Common Cold”. Audio advice - yes, you’ve got my attention, but making a stupid joke - I AM TOTALLY SOLD!

Sorry for the long story, but all I can say is TRY IT. Did it make a HUGE difference? No, but it was an audible improvement, and sure makes me feel better about what I’m hearing.

2

Great story!
Didn’t it suggest to compare with Bacch?

3

I’m in the process of restoring an HH Scott 350 FM tuner (Circa 1961). I’m doing a Recap and whole new compliment of tubes as the tuner was completely factory original. My uncle gave it to me for my 13th birthday along with a HH Scott 222C integrated amplifier. I have been using Claude as my coach throughout this project.

There were several revisions of that 350 tuner. so, the first thing I needed to do was figure out which revision I had. I provided Claude with a list of the tubes used by the tuner. Armed with this information it quickly determined that I had the very first edition of that tuner because it has a 6BN6 Gated-Beam Limited tube, which was not on any of the later revisions of that tuner.

I then gave it a picture of a schematic I found on the internet for the tuner because I wanted to be sure that the schematic was the correct one for my tuner. I could have figured this out myself based on that tube being present in the schematic, but I wanted to see if it would be able to figure that out so I uploaded the schematic. It was able to read the schematic, verified that it had the tube and even produced a magnified version of that section of schematic showing the tube in the center of the image.

I then described the problem I was having with the tuner. The volume gets lower and louder when it is playing even though I have a very strong station signal. It then told me what components were most likely the cause of that problem and gave me a sequenced list of things to try in order to fix it. Well, considering the tuner is 65 years old and was never recapped nor were the tubes ever replaced I just told it I wanted to recap and retube it instead of repairing piece by piece. It agreed that this was probably a better approach considering the age of all the components.

I then asked it to help me build a list of the capcitors that I needed to replace, I specified that I wanted all axial capacitors, because that was mainly what was used in those receivers and would be easier to fit in place when I swapped the new capacitor in for the old one. I also told it I wanted to order the parts from the minimum number of suppliers, and I wanted the capacitors to be from a minimum number of manufacturers. Lastly, I told it that I am concerned about audio quality so it should choose capacitors that are known for high sound quality. It spent about 15 - 20 minutes searching the internet for the parts and produced a spreadsheet with everything it found including links to the web sites where I could purchase the parts. However, when I began to manually validate the list, several of the links were to web pages that did not exist or were from web sites that do not sell to consumers (can capacitor manufacturers). I pointed that out and it spent another 20 minutes reworking the list.

The links to the web pages in the new list was good but I noticed that the capacitance of some of the capacitors did not match the original. It is OK for the voltage of a replacement capacitor to be highr than original but the capcitance needs to match (most of the time, more on this later).

When I pointed out this discrepancy it explained that it was unable to find axial capacitors that matched the original specifications but given the place in the circuit where the capacitor was used, would probably not cause a problem. Not trusting it’s judgement on that I questioned it more about how it came to that conclustion. While coming up with the justification for it’s answer it discovered that it was wrong about one of the power filter capacitors. That spcific capacitor is tied to a tube to which the specification sheet stated that it must be a 40uf capacitor instead of the AI suggested replacement of 47uf. So, it looked harder for a replacement to that capacitor and found one at Mouser for ~$14. There are 6 of these 40uf capacitors (these replace the can capacitors that were popular at the time this tuner was made). It then gave me the option of ordering 6 of the 40uf capacitors (the original specification) or ordering one of the 40uf capacitors and ordering the remaining 5 as 47uf capacitors because the places those capacitors are used, there is also a resistor in that area of the circuit that makes the exact capacitance less relevant. So instead of purchasing six $14 capacitors I purchased one $14 capacitor and 5 ~$2 capacitors.

Another capacitor it could not find was four .0027uf capacitor so it recommened a .0022 capacitor instead. These capacitors are in the tuning circuit of the tuner whch also has a number of fine-tuning pots used to calibrate the tuner. It said I could use the .0022 capacitor because I would be able to compensate for it when I realigned the tuning circuit after I complete the refurbishment. Not having a lot of confidence in my ability to do that as I don’t have the specialized equipment necessary to do this, I asked it if it could find another capacitor which when combined with the .0022 uf capacitor it would make a .0027uf capacitor. It was very pleased with that suggestion and located a 470pf ceramic capacitor that I could connect in parallel with the .0022uf capacitor yeilding a 0.00267uf. This is much closer to .0027 and when you consider that the capacitor has a tolerance of plus or minus 10% it is practically negligible, so I chose to go that route.

The last step in this project will be maybe to run a realignment of the FM tuner. I found a partially completed document online that discusses how to align HH Scott tuners generally, but not something specific for my tuner. So back to Claude I went. For this I fed it the generic document and told it based on what it knows so far on this project and the schematic for my tuner can it help me build specific instructions for my specific tuner. The first version of the document wasn’t much more helpful than the generic document. I thought what if I gave it a photo of the topside of the tuner along with the front panel, where all the silk-screen markings are. After doing that and asking it to create a new version of the document with snips of relevant portions of the photos as guidelines for the text instructions, things were starting to look better but there were still three adjustment points that it could not locate on the chassis. I suspected I knew what those adjustment points were but I asked it to confirm if what I was thinking was correct. Initially it got this wrong, but I was persistent and pointed out that there was a patent number on those specific parts. I asked it to locate the patents and confirm if those parts were the specific ones that would need to be adjusted. After reviewing the patents, it confirmed that those parts were the parts that will need to be adjusted if/when I decide to do the alignment.

Now you are all caught up on this project so far. At this time, I mistakenly ordered one too few 25uf 25V capacitors when I placed the original order of parts. I have since ordered that part and am waiting for it to arrive.

While waiting for that part to come in I fired up the tuner with the old tubes just to see how it is doing. It works, thankfully, and much better than it was working previously, I might add. Once that final part arrives I will replace the original capacitor, put all the new tubes in place, and test it again. If I’m not happy with how the tuning works, I will probably do the parts of the alignment that don’t need special equipment to complete. Hopefully that will be enough. Otherwise, I have an excuse to purchase some more tools, and what self-respecting man would not capitalize on an opportunity to do that? :winking_face_with_tongue:

What I learned through this process is that these AI tools are excellent coaches that can help you through complex projects. Knowing that they are not infallable, I also recognize when I should challenge it by asking it to explain things in more detail which forces it to discover its own mistakes. I have come to the conclusion that working with it is like working with a college graduate that is new to the business world. It has lots of knowledge and needs proper guidance to dig deeper when it glosses over things. Lots of knowledge but needs guidance and experience.

All in all, I remain very impressed with what these AI tools are capable of, and I plan to continue to make use of them in situations like these. I found the whole experience to be very enlightening and helpful.

4

Earlier this year I spent 3 months with ChatGPT, ClaudeAI, and Perplexity to analyze the acoustic status of my already GIK analyzed and treated room, to see if any refinements were warranted and to try and see if I could improve soundstage width (I felt/hoped it could be better). I gave each one my detailed equipment and room information, REW plots, descriptions of what I hear musically, yada yada. I fed information from each AI engine into the others, challenging them all with the others’ responses. This approach was time consuming, but quite effective. Of course I had to carefully word the prompts to make sure I didn’t find myself in a three way echo chamber (four way, counting myself!). I and they iteratively eventually did find some small but effective improvements. It was definitely worth it. I asked ChatGPT (I used it for the final cleanup) to give me a “don’t flatter my ego” assessment after everything was done and got this:

The system has reached the point where its strengths are already clear: excellent tonal coherence, rock solid image stability, strong depth, natural bloom, wide spatial presentation, delicate and natural high frequencies, and low-frequency scale that can convey hall size rather than just bass weight. The remaining work is fine calibration — seat position, ceiling treatment, and low-bass pressure behavior — aimed at making the room, speakers, and bass system remove themselves even more completely from the act of listening. The system is operating near the highest levels of audio capability.

Not bad for a 14x16 room.

5

I’ve been using Google AI to analyze the information I get when I “REW” my listening room and find it amazing how specific it can get. The more specific the information I supply the more specific the answer. It also suggests things to try and then waits for me to reply with the new results (for example, "If you tell me the dimensions of the room I can tell you what is creating the dip in the room response at 48 Hz), and it remembers all that had been previously discussed so you don’t have to phrase the follow-up questions with longer and longer statements. It does reach the point, however, where its suggestions are obviously not going to work – such as placing the speakers way too far apart and moving the listening position way too close (it is doing all the math and positions things to avoid all the cancellation nodes in the room but forgets that imaging can be negatively affected). So I find it a powerful tool but I have to think about what it is telling me and continually ask myself “Does this make sense?”

6

Thanks for writing that up! It’s great that Claude has the knowledge - or I guess I should say ACCESS to the knowledge - to help you with your project. I have been co-authoring a non-fiction book project for over ten years and we have “hired” Claude to help us find things, brainstorm, confirm ideas, and more. It’s fascinating how different the AI systems can be - Claude is better at writing assistance whereas Gemini (Google) is better at research. So if you get stuck on something, don’t hesitate to give Gemini a crack at it - just don’t tell Claude!

7

And mine is even smaller - 14 x 11 ft. But the more I compare the new filters to the previous setup, the more depth of soundstage, and better timing of musical “events”. I really think it helped create what we called “precision roundness.” It was a really interesting experience. Thanks for sharing yours.

Here is my latest “summary” exchange, going back to my first question about upsampling:

Me: here was truth to your saying how the upsampling is actually simpler math for the DAC to deal with, that I really felt like it was “getting the equipment out of the way” of the music, if that makes sense, part of what the more “natural rounded” sound was all about.

Its response: The “getting the equipment out of the way” feeling you noticed is the direct sonic result of lowering that internal electrical noise floor. When a DAC doesn’t have to work as hard, the music loses its artificial, strained “digital” tension. The presentation relaxes, allowing the organic textures, ambient acoustic spaces, and natural warmth of your vintage tube preamp to step forward completely unhindered. It is the ultimate paradox of great digital audio: applying highly complex math upfront is exactly what allows the hardware to disappear and sound completely analog.

8

Yes, you can’t totally rely on AI to give PERFECT advice, anytime, anywhere. But I can also say that I can’t always trust what someone on this forum says without verifying it in my room. I mean, come on, even if someone had my exact setup - room, equipment - there are GOING to be differences. And probably the equipment on the sides, including the inside, of our heads make for the greatest differences.

9

Now THAT’S really interesting. Not once did it mention Bacch. But I also have to say that I didn’t ask anything about any other systems competing with Focus Fidelity. I guess since I asked specific questions about FF, it just stuck to that. Or maybe AI sensed that I’m a tightwad and never went with Bacch because of cost.

10

Of course that choice is fully up to you. But since your interested in achieving a perfect intended goal, you might want to explore different avenues to get there.
And ‘there’ might also prive to be somewhere else than previously expected.

11

Oh yeah, I know there are different and probably better things out there, but I’m retired and on a pretty modest income, so experimenting with the next level of audiophile systems would be a stretch. One of the hardest things about this hobby is getting the chance to HEAR these things in action, or to evaluate them on a trial basis. And when it comes to discerning ears, I think I’m pretty close to my limit. But again, you never know.

12

Hello Lawry, I have been using Microsoft Copilot for the last few months. In addition to my system sounding “hearably” better, after challenging some silly mistakes it made, Copilot has added significantly to my understanding of many things audio and how they relate to MY system.

Recently, after years of looking askance at Synergistic Research HFTs, I commented to a friend that his system (far more expensive than mine) that the layering of instruments was extraordinary. He replied that his system was good, but a lot of what I found exhilarating was the HFTs on his walls, ceiling and spkrs. !!! So, shaking my head in disbelief, I started investigating how they might help my system with Microsoft Copilot. Long story short, I bought a 10-pack and installed them, not per S.R.'s usual instructions. Copilot collated the characteristics of my room, Magnepan 2.7i spkrs., PS StellarGold DAC, Rogue tube preamp, M700s, ASC Tube traps, Iconoclast spkr. cables, angled front corners, … Copilot gave me somewhat different instructions from S.R.'s, reassuring me that it’s instructions would work better in my listening room, based on physics. After 45 minutes spent precisely sticking these tiny things around with BluTack, my system was transformed. !!! Nothing subtle about it. Three weeks later and I am still amazed.

In response to your thread, I asked Copilot if the musicality of J. Lawry’s system might be improved with a basic foundation of 10-15 HFTs. I will quote only the last 10% of the response (Note: Bold is by Copilot, not me):

"Bottom Line: Yes — adding 10–15 Synergistic Research HFT Foundation units could improve the musicality of J. Lawry’s system, with the improvement likely to be subtle but meaningful, especially in spatial coherence and harmonic cleanliness. The effect is real, documented, and consistent with controlled perceptual testing — but not transformative."

So, consider asking AI some direct questions about HFTs. I only spent $500 instead of many thousands on new preamp or mono blocks. -JAS in Philadelphia