• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Could ChatGPT Replace Audio Writers?

sonitus mirus

Active Member
Joined
Apr 19, 2021
Messages
256
Likes
336
Be sure to post the coding error to chatGPT, as it can oftentimes make the correction. Sometimes it makes me lazy, or is it saving me time for other things?
 

Axo1989

Major Contributor
Joined
Jan 9, 2022
Messages
2,820
Likes
2,816
Location
Sydney
We don't cry foul when a spell checker is used. In time, this would be accept just the same.

Well, I cry foul quite often as iOS spell-checking gets increasingly creative with word substitution. It's not unlike the AI-generated texts posted in this thread (including your bogus biography example) in that looks like something I might have written, but it's just wrong.

Arstechnica comment threads following articles on ChatGPT and similar subjects have been pretty interesting. One article I didn't expect: ChatGPT-style search represents a 10x cost increase for Google, Microsoft. Whether those estimates hold up will also be interesting.
 

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,632
Likes
10,207
Location
North-East
I would not give up on ChatGPT @amirm but like anything new will take a little bit tweaking both peoples expectations of what it can do and I believe how say OpenAI (and now Bing) and maybe ultimately Google tweak their algorithms. I am not sure if people are aware but the API provided by OpenAI has a few parameters that can be tweaked.

Specifically there is one called "temperature" and when using the API varying this value controls how "creative" ChatGPT (GPT-3 in the case of the API) is. I believe the current ChatGPT is something like 0.5 - making this 0 means that it does not create "random" content and sticks with "completing" text based on what it has seen. So I believe ultimately the UI might have some simple tweaks / options e.g. I am writing an essay therefore please be a little creative or I need a factual response.

For what I do (I am a software architect / engineer) - I find it astonishing good (I presume people know it can generate "code" in various software languages. So for me leaving it be a little "creative" has been interesting - sort of like working with a Junior Engineer (they get some of the specifics wrong but I can take there work and update it (refactor it) into something useful. But I know with software when it get it wrong "creatively" it becomes obvious fairly quickly e.g. it can make up the names of functions / method within other systems that don't exist so ultimately I get an "error". but as a aid for rapid development of mostly good structured code I would say its as an amazing tool.

Hah! I asked it to give me an example of how to use a specific audio library to output DSD.

ChatGPT obliged with some great looking code, well commented... and completely wrong. It had made up a bunch of objects and functions that were never part of the library, gave them reasonable and understandable names and overall a correct structure... except it would never work because none of them were real. I tried to find where these made-up objects and functions came from, but was unable to locate even a single example in any of the source code repositories or through Google search. My guess is that it was a mash-up of some real examples with some code stolen, possibly from some private repositories by ChatGPT. In any case, completely useless, but great looking ;)
 

krabapple

Major Contributor
Forum Donor
Joined
Apr 15, 2016
Messages
3,169
Likes
3,717

sarumbear

Master Contributor
Forum Donor
Joined
Aug 15, 2020
Messages
7,604
Likes
7,314
Location
UK
It is amazing how detailed the false information is, making one believe that it has to be right.
That is the problem with the current LLM models.
 

Spocko

Major Contributor
Forum Donor
Joined
Sep 27, 2019
Messages
1,621
Likes
2,999
Location
Southern California
One of my colleagues challenged Chat GPT to explain the Fletcher-Munson curve, here's the result.

My Question:
In the context of mastering how can the Fletcher Munson equal loudness curve be applied to achieve a consistent perceived level by the listener.

Chat GPT:
The Fletcher-Munson equal loudness curve is a graph that shows the relationship between sound pressure level (loudness) and frequency, at different loudness levels.

The curve illustrates how our ears perceive different frequencies to be at different loudnesses at different overall loudness levels.

When mastering audio, this curve can be used to ensure that a consistent perceived loudness is achieved across different frequencies and loudness levels.

This can be done by adjusting the levels of different frequency bands so that they align with the curve, resulting in a balanced and consistent perceived loudness for the listener.

My Question:
That was a good answer, could you expand on the type of equipment you would use to a balanced and consistent perceived loudness for the listener.

Chat GPT:
In order to apply the Fletcher-Munson equal loudness curve during mastering and achieve a balanced and consistent perceived loudness, a mastering engineer would typically use a combination of equalization (EQ) and dynamic range compression.

EQ is used to adjust the levels of different frequency bands in the audio. By boosting or cutting certain frequencies based on the Fletcher-Munson curve, the engineer can ensure that the audio has a consistent perceived loudness across all frequencies. A common EQ technique is to use a multiband compressor or a parametric EQ.

Dynamic range compression is used to control the difference between the loudest and quietest parts of the audio. By reducing the dynamic range, the engineer can make the overall loudness of the audio more consistent. Compressors, like FET and VCA compressors, are the most common tools used for this purpose.

Additionally, a mastering engineer might also use a loudness meter to measure the perceived loudness of the audio and ensure that it falls within a desired range. This will give a numerical representation of the loudness, and the engineer can adjust the levels accordingly.

It's worth noting that while these are common techniques used to achieve a balanced and consistent perceived loudness, the process of mastering audio is not an exact science and mastering engineers may use other techniques or equipment to achieve the desired results.
----------------------------------------------------------------------------------------
Not bad eh?
DANG, I'm waiting for the day I feed it an audio mix and ask it to suggest EQ level adjustments to clean it up based on certain assumptions and best practices standards - we're almost there at this rate
 

Spocko

Major Contributor
Forum Donor
Joined
Sep 27, 2019
Messages
1,621
Likes
2,999
Location
Southern California
...
Overall, while human product reviewers bring unique perspectives, ChatGPT as an AI language model can complement and enhance their work, providing a fast and reliable analysis of audio products.
"unique perspectives" in my mind also include the individual yet concealed preferences and foiables that are not necessarily relatable because they arise as a result of:
  • hearing loss (age related or damaged)
  • cognitive spectrum accommodations (ADHD invididuals have incredibly strong biases for/against certain sound signatures and frequencies)
  • listening volume (the music listening experience that can be gleaned at 73dB is much different than at 83dB)
 

Axo1989

Major Contributor
Joined
Jan 9, 2022
Messages
2,820
Likes
2,816
Location
Sydney
Archimago has audiophile questions for ChatGPT


Interesting reading. You can see how ChatGPT delivers a low-temperature result compared to the initial version of AI Bing search, which was prompted to output some crazy hallucinations. The downside is it's all bit bland and basic.

I did expect something more incisive from Archi though, based on his usual writing. Perhaps his (imho) overly positive response is based on some underlying analytical assumptions: he says "pretty evident that based on what it knows" apparently without considering that ChatGPT doesn't know anything, it constructs elaborate text strings built on the most likely next word in any string, based on the large language corpus and the input query/prompt. That text is serviceable (as final result or input) for many purposes, of course.

* "temperature" and "hallucinations" being terms of art
 
Last edited:

sarumbear

Master Contributor
Forum Donor
Joined
Aug 15, 2020
Messages
7,604
Likes
7,314
Location
UK
DANG, I'm waiting for the day I feed it an audio mix and ask it to suggest EQ level adjustments to clean it up based on certain assumptions and best practices standards - we're almost there at this rate
The ChatGBT answer is completely wrong! It doesn’t understand the concept of loudness and confuses it with mixing.
 

Axo1989

Major Contributor
Joined
Jan 9, 2022
Messages
2,820
Likes
2,816
Location
Sydney
The ChatGBT answer is completely wrong! It doesn’t understand the concept of loudness and confuses it with mixing.

Yes, as you implied earlier, generating responses via probability based on a large language corpus is orthogonal to understanding and delivering a correct answer.
 

Spocko

Major Contributor
Forum Donor
Joined
Sep 27, 2019
Messages
1,621
Likes
2,999
Location
Southern California
The ChatGBT answer is completely wrong! It doesn’t understand the concept of loudness and confuses it with mixing.
the big question is whether it will improve? I know garbage in garbage out, but who's in charge of correcting the garbage?
 

LTig

Master Contributor
Forum Donor
Joined
Feb 27, 2019
Messages
5,760
Likes
9,442
Location
Europe
It is amazing how detailed the false information is, making one believe that it has to be right.
Absolutely. I asked it about our NMR software and its answer was in parts totally wrong but it was given in such a straight faced manner that someone who is not familiar with the subject will believe it to be true. Hence I see big problems coming when people ask AI questions about subjects they have not enough knowledge to be able to check the validity of its answers.
 

Chrispy

Master Contributor
Forum Donor
Joined
Feb 7, 2020
Messages
7,757
Likes
5,918
Location
PNW
lol you mean more entertaining, more informative or more articulate? As a creator of product review videos myself, I'm definitely curious as to your thoughts on this!
Yes. I find most subjective reviews only worthwhile as entertainment, tho....and many youtube type reviews suck.
 

skraz

Member
Joined
Jul 6, 2022
Messages
16
Likes
27
Absolutely. I asked it about our NMR software and its answer was in parts totally wrong but it was given in such a straight faced manner that someone who is not familiar with the subject will believe it to be true. Hence I see big problems coming when people ask AI questions about subjects they have not enough knowledge to be able to check the validity of its answers.
With misinformation such a problem in the world as it already is, do you see this as being even worse in regards to spreading false information going forward?
 

Axo1989

Major Contributor
Joined
Jan 9, 2022
Messages
2,820
Likes
2,816
Location
Sydney
the big question is whether it will improve? I know garbage in garbage out, but who's in charge of correcting the garbage?

One answer to that question is that the LLM (large language model) can't practically be curated for truth by humans, the corpus is too vast. Replace "large" with another size and it might be possible (build a model based on Wikipedia perhaps, but there will be language and subject gaps, obviously).

If you want accurate answers, rather than probabilistic text, another modelling approach is required. Once you start adding rules, the problem space obviously increases. In replicating what a human might say, the ChatGPT model does a very rapid and somewhat reasonable job if you don't mind slogging through an uncanny valley of blandness. But in feeding a model everything we can find that humans have written we provide that the output will include errors and lies, because we've written plenty of them. There will be errors (as we've seen here) largely because the model doesn't know anything. Calling them lies is anthropomorphising of course, because the model has no motives, it isn't sentient or sapient (except it in the sense that it may appear so).

Another answer is that we are in charge of correcting the garbage, by testing these AIs to breaking point over and over. Ideally, the model-makers then build better models. And we'll use them to build better worlds. :)

 
Last edited:

HarmonicTHD

Major Contributor
Forum Donor
Joined
Mar 18, 2022
Messages
3,326
Likes
4,830
So far I have had good experiences in my queries to ChatGPT. That ended when I asked it to tell me who I am (by name). The start was just the regurgitation of LinkedIn profile or maybe my bio online. But then the second paragraph had this total nonsense:

After leaving Microsoft in 2009, Majidimehr co-founded a startup called StorSimple, which developed a cloud storage solution for businesses. The company was acquired by Microsoft in 2012, and Majidimehr returned to the company as a corporate vice president for the Cloud and Enterprise division. He left Microsoft again in 2015 to join Verkada, a startup that develops cloud-based security systems.

Other than the date I left Microsoft, none of this is remotely true. I searched for my name associated with StorSimple and google essentially gives up, not finding any connection. I also searched against my first name in connection with that company and while I got more hits, none made much sense as the foundation for the above.

It is amazing how detailed the false information is, making one believe that it has to be right.
:)

So Microsoft still owns you money for the company they bought from you. The proof is their own software ;)
 

sarumbear

Master Contributor
Forum Donor
Joined
Aug 15, 2020
Messages
7,604
Likes
7,314
Location
UK

FrantzM

Major Contributor
Forum Donor
Joined
Mar 12, 2016
Messages
4,340
Likes
7,742
Hi

I may be stuck on repeat :D:

We seem to be amused and comforted by the fact that these AI make mistake, returning some times errors or nonsense... That is how we learn. We make mistakes, often, repeatedly, then we know, then we act...
This thing can become very dangerous. Powerful entities with the potential of absolute control over any human. Think about it. We already have computers controlling, virtually all aspects of our lives.... and, People! They are connected!! Yes they are. They talk to each other all the time... Right this moment , we the Humans seem to be in control.. We are however, happily, gleefully even, abdicating this control... to something that we don't know or understand very much. What stops AI to lie? To control elections results? The Markets? Our lives?
Scared.

Peace.
 
Top Bottom