GPT4 / CHATGPT Plugins for Speaker measurement analysis and comparison

[Curvature]

Balls in a circle aside, any progress in feeding GPT measurement data and asking for results?

 

[Jeromeof]

Balls in a circle aside, any progress in feeding GPT measurement data and asking for results?

Nothing from me yet - while I believe the plugin approach will solve this ultimately - the interface between GPT4 and the plugin's is very buggy at the moment and I believe some of the recent (sort of boring) changes that OpenAI are making (e.g. defining a JSON schema for the chat to follow) are about fixing this core problem first.

Q: Calculate a closed speaker box's volume to have a Q of 0.71, using a driver with the following parameters: Fs: 20Hz, Vas: 113l, Qes: 0.31, Qms: 5.1

I gave your problem a go last night with GPT4 (and then with the Wolfram alpha plugin) and I got it to generate the correct python code but only after I corrected the core formula. I am still in a 'holding pattern' as far as using chatGPT / GPT4 for useful science. But what your problem did give me was possibly an idea for an audio science specific plugin which when included would return the correct formula for a given term, as I don't think basic chatGPT or even Wolfram has all the specific terms and algorithm's required.

This would mean the solution would be a mixture of 3 or 4 things:
1. GPT4 - translate English into 'python code' (I am still hopeful of their 'code interpreter' plugin)
2. A plugin to provide clear definitions when required - in this case return the formula to feed into the python code
3. A source of raw data - e.g. spinorama (and possible ASR for amp / dac ) data
4. Something to run the generated code - The 'code interpreter' and ultimately provide visualisations and other tools e.g. export as CSV

 

[F1308]

I think you missed my last post.

Oh,well...finally !!!!

 

[F1308]

Balls in a circle aside, any progress in feeding GPT measurement data and asking for results?

Impossible while BALLS AROUND A ROD gets a wrong number.

 

[sarumbear]

Nothing from me yet - while I believe the plugin approach will solve this ultimately - the interface between GPT4 and the plugin's is very buggy at the moment and I believe some of the recent (sort of boring) changes that OpenAI are making (e.g. defining a JSON schema for the chat to follow) are about fixing this core problem first.

I gave your problem a go last night with GPT4 (and then with the Wolfram alpha plugin) and I got it to generate the correct python code but only after I corrected the core formula. I am still in a 'holding pattern' as far as using chatGPT / GPT4 for useful science. But what your problem did give me was possibly an idea for an audio science specific plugin which when included would return the correct formula for a given term, as I don't think basic chatGPT or even Wolfram has all the specific terms and algorithm's required.

This would mean the solution would be a mixture of 3 or 4 things:
1. GPT4 - translate English into 'python code' (I am still hopeful of their 'code interpreter' plugin)
2. A plugin to provide clear definitions when required - in this case return the formula to feed into the python code
3. A source of raw data - e.g. spinorama (and possible ASR for amp / dac ) data
4. Something to run the generated code - The 'code interpreter' and ultimately provide visualisations and other tools e.g. export as CSV

Have you checked Google's Bard?

 

[sarumbear]

Oh,well...finally !!!!

You may say so, but you were the only person who went to trouble to show me example of ChatGPT failing while using speaker calculations. I thank you for that. Everyone else was either just saying it won't work or showing unrelated examples.

 

[Jeromeof]

Have you checked Google's Bard?

No because it seems like the EU and Google are not getting on so Bard is blocked in Europe

 

[sarumbear]

No because it seems like the EU and Google are not getting on so Bard is blocked in Europe

The operative word "it seems" is never a good way to know things

I'm in the UK and I can use it. I VPN'ed to a German IP and I can still connect to Bard.

 

[Jeromeof]

The operative word "it seems" is never a good way to know things

I'm in the UK and I can use it. I VPN'ed to a German IP and I can still connect to Bard.

That is because the UK is not in the EU and google is actually pretty good (or bad depending on your POV) about detecting your real location even when you use a VPN.

 

[sarumbear]

That is because the UK is not in the EU and google is actually pretty good (or bad depending on your POV) about detecting your real location even when you use a VPN.

It looks like you can use Bart in the EU if your account was created outside the EU. My account was originally created in the UK so that may be why I can use it, not because of IP location.

 

[F1308]

You may say so, but you were the only person who went to trouble to show me example of ChatGPT failing while using speaker calculations. I thank you for that. Everyone else was either just saying it won't work or showing unrelated examples.

Regarding volumes...


Sphere Radius: X Calculation


User
I computed the surface of a sphere and it was X m2.
Then I went computing its volume and it was X m3.
What is the radius of that sphere ?

ChatGPT
To find the radius of a sphere given its surface area and volume, we can use the following formulas:

Surface Area of a sphere:
A = 4πr²

Volume of a sphere:
V = (4/3)πr³

Since you have already computed the surface area and volume and both are equal to X, we can equate the equations:

4πr² = X (Equation 1)
(4/3)πr³ = X (Equation 2)

To find the radius, we can rearrange Equation 1:

4πr² = X
r² = X / (4π)
r = √(X / (4π))

So the radius of the sphere is √(X / (4π)).





User
I need the radius in m. I need a number.

ChatGPT
Apologies for the oversight. To provide you with a specific numerical value for the radius, I would need the actual value of X (surface area and volume) in meters squared (m²) and meters cubed (m³) respectively. Could you please provide the specific values you obtained for X in order to calculate the radius accurately?





User
The trick here is that we don't need a number as it turns to be the same for volume and for surface, as already stated.

ChatGPT
Apologies for the confusion. If the surface area and volume of the sphere are indeed equal (X m² and X m³ respectively), it implies that the radius of the sphere is the same for both calculations.

Since we know that the formulas for surface area and volume are:

Surface Area of a sphere:
A = 4πr²

Volume of a sphere:
V = (4/3)πr³

And in this case, A = V = X, we can equate the equations:

4πr² = X (Equation 1)
(4/3)πr³ = X (Equation 2)

Since X is the same in both equations, we can conclude that the radius of the sphere is the same value for both calculations. However, without knowing the specific value of X, we cannot determine the exact numerical value of the radius.

 

[F1308]

That means AI does not know how to
resolve r
when
X=4*Pi*r^2
and
X=4/3*Pi*r^3
or
4*Pi*r^2=4/3*Pi*r^3

A sphere with same values for surface and volume has a radius of 3 units.


And yes, when asked for r in 4*Pi*r^2=4/3*Pi*r^3
the correct answer is given, 3, but we were talking about intelligence.

Equation solver did it first...

https://play.google.com/store/apps/details?id=com.alago.eqnlite

 

[Jeromeof]

So finally (after months of waiting) OpenAI released the 'Code' plugin - its early days but here is a preview of what it can do.

So I decided as a test to see could I get a nice graph comparing the on-axis response of 2 speakers from ASR data (stored on spinorama).

For now (until the spinorama plugin is ready), I just download 2 of the CEA2034.txt files and then did the following:

It then struggled to understand the format - I think it really likes CSV files rather than this format - but it persisted - it tried various ways until it loaded the data from both files but it still struggled to understand which column was which dataset. I think ultimately the spinorama plugin might need to translate the CEA2034 dataset into sometime more suitable, but for now I am giving this raw data.

It failed to understand the following:

But then I clarified (to make it easier for now) that I only wanted the first column, the 'On Axis' column compared

Obviously the y axis scale is wrong so I suggested a different Axis:

Is this useful on its own? - probably not, but the key point is that each stage of this process is actually 'code' so can be checked for accuracy e.g. the "Show work" drop down looks like this:

And can be made more complex by just asking it in English. e.g. 2 simple expressions to manipulate the code which changes the chart:

And then:

I then decided to add another speaker to the chart:

Its very early days (I think they only released access a few hours ago) so I will play around a bit more tomorrow but I hope this explains a little why I see this plugin as the most useful for audio science (assuming we can get the various raw data into a good format for analysis).

 

[Jeromeof]

So I had a reason to use a little more of the chatGPT 'code' plugin in the last few days. I thought I would create a post about the FiiO FX15 (as FiiO send me some raw measurements) - that new post is here so while I could use REW for the Frequency Response graphs and some comparison graphs I thought I would generate some extra graphs for the post. I am sure there are better tools but as an experiment I thought why not try and see what it could do for to help me generate some more graphics.

The RAW measurements spreadsheet provided had a Phase Response tab so I thought I would load the Excel spreadsheet into the ChatGPT "code" plugin - and effectively just ask it to generate a graph for me:

More interesting for me was I thought for the post I was creating I should try to match some of the graphs with similar graphs that @amirm would normally have in his reviews. The Phase graph was therefore not enough. So after looking up the formula (I didn't want to take a chance on a 'hallucinated formula' ), I ask it the following:

So I thought both graphs were good enough for my small post about the IEM, obviously if this was a term paper or some official document I would be double and triple checking what was done, the 'code' UI does provide a way to check the code generated (and I repeat this is not chatGPT generating data, chatGPT is generating code which does the analytics on the data and then the code to render the result).

Here is a view of the 'code' it generated in this case:

Anyway, while this is not an extensive example I found this incredible useful and probably saved me a few hours finding an alternative tool and transforming the data I had by hand to generate a chart that I could use for the post.