You might try sending ATE0 to the terminal to turn off the echo.
One thing often overlooked in the realm of AI is how the way a large language model operates will affect the responses you get. In a somewhat niche community like this there is not as much data to pull from, so like resolve said, most of the results are just what the most people are talking about. AI can’t subjectively analyze sound and attribute descriptors to it like people can, so when it describes headphones as bright, dark, full, etc. It’s just regurgitating someone’s opinion about the sound profile with no way to know if it’s accurate or not. It’s probably a safe bet to say that it is accurate, but still the fact remains the AI cannot determine that.
Maybe AI will endeavor to create bespoke, random or (yuck, shuffle) playlists - which allow for secondary and maybe even tertiary level preferences in what plays when. Off topic re sound quality, but equally or, in my case, more important since my sound is already pretty damned amazing.
The problem with automated recommendations is they don’t know why you like the piece and therefore have no basis for determining similar music.
Yes, they can assign their own generic attributes (orchestral, vocal, rap) but that’s just matching on their preferences, not yours. They can also do “other people who like this piece also like these other ones” and throw those at you.
You won’t get truly personalized recommendations until you can supply your reasons for liking the peice.
Here’s one thing AI could help with that more than just audiophiles could care about: make better hearing aids that consistently improve speech intelligibility, not just audibility.
(Also, some “good distortion” is a part of normal hearing? WTF? 
)
When a sound with two frequencies enters the ear, an additional sound is created by the cochlea itself at a third frequency that is a complex combination of the original two. These distortions are, of course, what we measure as distortion product otoacoustic emissions (DPOAEs), and their absence indicates impaired cochlear function.
But these distortions aren’t only transmitted out of the cochlea into the ear canal. They also elicit neural activity that is sent to the brain. While a hearing aid may restore sensitivity to the two original frequencies by amplifying them, it does not create the distortions and, thus, does not elicit the neural activity that would have accompanied the distortions before hearing loss.
These distortions themselves may not be relevant when listening to broadband sounds like speech, but they are representative of the complex functionality that hearing aids fail to restore. Without this functionality, the neural activity patterns elicited by speech are very different from those that the brain has learned to expect. Because the brain does not recognize these new patterns, perception is impaired.
[…]
But there is reason for optimism. In recent years, advances in machine learning have been used to transform many technologies, including medical devices (Nature . 2015 May 28;521(7553):436). In general, machine learning is used to identify statistical dependencies in complex data. In the context of hearing aids, it could be used to develop new sound transformations based on comparisons of neural activity before and after hearing loss.
Of course it’s more of a reference of what others say, since as a Large Language Model that comprises of most Generative AI; it’s going to just regurgitate what has already been said.
I don’t (or at least I hope) think that AI will change the landscape of headphones, as I feel the human element of understanding technology as a whole is so much more valuable. There’s so much GenAI does that kinda removes a lot of interaction with others, and I don’t value the fact there’s little human element to AI. I don’t think an AI truly has any reason or understanding of a person like what a real human can have. People who just reference others’ opinions on audio instead of truly forming their own are often looked down upon in most of these circles, because it begins to indicate they have no idea what they’re talking about. This is basically what GenAI is doing. Which is precisely why I do not trust it. It won’t really be able to understand the nuanced philosophies and values a headphone enthusiast would be looking for, and I will always believe an expert who shares similar values will have a much more conclusive and helpful opinion.
The human interaction is what I find to be very important too. I would not be here today if it wasn’t for reviewers, discussing things with aspiring reviewers like Listener who I am so proud to see he’s part of your community, and just interacting with people. Obviously it’s not optimal, and sometimes you get bad advice, but I just found the journey a lot more satisfying than just asking a LLM what to buy, and likely resulting in the same experience, but nothing in the way of social stimulation and connection.
There’s just too many problems that I believe will always be constant with Generative AI, and the lack of humanity involved with it makes it hard to consider of value.
2 Likes
I hate the term AI. I’m not sure that existing Generative models will have much impact, their better at summarizing things, extracting facts from documents and generating boilerplate than they are at design or reasoning tasks.
I could see them in use for say adjusting EQ based on user descriptions of what they hear, but I don’t think there is a lot of applicability much beyond that.
I suppose there is probably a place for it in audio signal generation, possibly in “correction”.
I think ML generally has a place, and it’s likely been in use by a lot of companies already.
All classical ML algorithms are just function approximators, they let you build an approximation of a function without having to know what the function is, just some set of inputs and outputs.
I’ve always thought that there was a real potential to start to quantify how subjective terms map to measurements, I’ve said this before but just because someone describes a particular sound as “Bass Light”, it doesn’t mean from a measurement standpoint it lacks bass, and that’s before you get into airy, spacious, open and other ambiguous terms. The idea is simple enough, you take a training set with subjective term, and measurements, more that just a frequency sweep, you then try and classify your test set to produce subjective terms from measurements with the model. Assuming it shows solid classification, then you train models with various input parameters missing, and see how the quality of classifications change.
There’s more to it than that subjective terms aren’t used consistently, but that really shouldn’t be a blocker, since I suspect there is a lot of commonality in subjective descriptions. And if there isn’t that’s a finding as well. ML is really all about feature selection so it’s not trivial but it’s something I wanted to try for a long time, the biggest issue being building that labelled dataset.
You could also use the model to potentially predict how a piece of audio equipment might be subjectively described from it’s measurements, so there is probably value in someone doing it.
Another potentially interesting idea is to classify users by their subjective descriptions of equipment, and preferences, and use that to suggest how they may like or not new equipment, this is just the pandora relevance model, users like you think this of X, so you might too.
And both of those ideas could have been done 20 years ago, so I’m not holding my breath.
2 Likes
Excited to realize a new use for ChatGPT: songs lyrics interpretation. I was having trouble understanding what the song Dissipating by Wheel is all about. First I tested ChatGPT with Lateralus by Tool. The result wasn’t just right, it was impressively comprehensive. The results for Dissipation gave me a new level of appreciation for that song and Wheel in general.
1 Like
I tried Google AI with a random selection of bands, and somehow ended up with The Flaming Lips. With an obvious contrast to the prior reviews, the AI positively gushed and praised them to the heavens. So special, so tender, so unique. I responded by saying “they are just a band” – the AI of course capitulated. Training sources matter.
First-pass AI answers tend to be popularity contests, and if you know a domain you can sometimes tear them apart. Still, a lot of AI is Wikipedia with an interactive interface.
1 Like
Yep. LLM AI is very good at summarizing a single source of truth or when there is consensus. Not so good at opinions since there is rarely consensus.
I prefer asking ChatGPT for information over Google search or Wikipedia. The other day I used ChatGPT instead the car manual because GhatGPT is actually more specific to the year and trim level.
1 Like