Our perception of audio

[andreasmaaan]

I do think though that this book analogy is a bit of a stretch.

The question whether an audio system’s non-ideal transfer function might be preferred by listeners over the transfer function of a (closer to) ideal system is best settled empirically, IMHO.

 

[Cosmik]

I’m sure AI or algorithm based editing of all kinds of written texts - eventually perhaps even the writing of them - will in fact be common (and best) practice in the near future.

A whole discussion in itself! Top of the Google list:
Racist, Sexist AI Could Be A Bigger Problem Than Lost Jobs
https://www.forbes.com/sites/parmyo...ial-intelligence-ai-bias-google/#32bb5ebf1a01

How about this as an argument? The necessity to paper over uncomfortable truths in the data will ultimately kill off the usefulness of AI.

In order to make AI 'PC', it is necessary for humans to tinker with the data and algorithms subjectively, and this *must* render it useless.

 

[Sal1950]

Sounds like you guys are discussing MQA

 

[Cosmik]

I do think though that this book analogy is a bit of a stretch.

The question whether an audio system’s non-ideal transfer function might be preferred by listeners over the transfer function of a (closer to) ideal system is best settled empirically, IMHO.

But what has a 'transfer function' got to do with sound, acoustics, music? Absolutely nothing. What knowledge does the transfer function have about what the signal contains or represents? Ditto.

The results it gives have nothing to do with the original piece of music; they are merely spurious, meaningless 'patterns' that humans may like as a novelty or superficial effect. Nothing wrong with that, of course, but it's just using an audio system as a dumb pattern generator seeded by the recording.

Like a child's kaleidoscope uses a bit of the real world to 'seed' its patterns. It's fun for about ten minutes.

 

[Blumlein 88]

Sounds like you guys are discussing MQA

Well more like whether there are euphonic systems a good many people might prefer to true accuracy on a reliable basis over lots of musical input. MQA I suppose could be considered a lame DSP based attempt at that. We both know it really was about audiophile myths turned into reasons for DRM.

 

[Theo]

Like 'replacing long words with shorter ones' would make a book easier to read?

Isn't that called "texting"?

 

[Theo]

"I've developed an algorithm to improve books"
"I've developed an algorithm to improve recordings"
To me, these statements are equivalent in terms of their credibility.

https://thenextweb.com/artificial-i...beatles-inspired-song-written-entirely-by-ai/
Not sure I would call that an improvement...

 

[svart-hvitt]

@Cosmik ,

does our nonlinear hearing mean the volume knob is a tweak? Should all audio playback be done at one specific preset level?

 

[andreasmaaan]

A whole discussion in itself! Top of the Google list:
Racist, Sexist AI Could Be A Bigger Problem Than Lost Jobs
https://www.forbes.com/sites/parmyo...ial-intelligence-ai-bias-google/#32bb5ebf1a01

How about this as an argument? The necessity to paper over uncomfortable truths in the data will ultimately kill off the usefulness of AI.

In order to make AI 'PC', it is necessary for humans to tinker with the data and algorithms subjectively, and this *must* render it useless.

I read that article very differently. The AI in question was programmed using predominantly white faces. It therefore became good at recognising white faces, but did not become good at recognising faces of people with darker complexions. The problem was then rectified by expanding the data set. A poor-performing system was made better.

What’s this got to do with political correctness?

 

[Cosmik]

@Cosmik ,

does our nonlinear hearing mean the volume knob is a tweak? Should all audio playback be done at one specific preset level?

If the recording was 'purist' then you could justify an argument that said that playback should always be at the same absolute level as the original.

Peter Walker of Quad said

"there is only one correct volume level for any particular piece of music."

 

[Cosmik]

What’s this got to do with political correctness?

I think it would be naive to think that AI will be allowed to produce politically incorrect results, and also naive to think that political correctness is always scientific correctness.

Creative AI will *always* need to be tweaked, vetted and curated by humans, or restricted to the anodyne. It will never even give you a slightly (these days) risqué title like 'Black Betty' or 'White Punks on Dope' or even 'Devil Woman'.

Cliff Richard's work will always remain more exciting than AI-generated books or music.

 

[andreasmaaan]

Is anyone aware of research into our brains' ability to compensate for equal loudness curves. It's well-established that the curves are basically correct and do not vary substantially from person to person (at least not in general trend), but given these curves are ever-present, it would seem surprising to me if we did not unconsciously compensate for them to a significant extent anyway.

I think it would be naive to think that AI will be allowed to produce politically incorrect results, and also naive to think that political correctness is always scientific correctness.

I agree (although it's debatable how much of an effect "political correctness" will have). And by the same token, it would be very naive to think that AI won't also be affected by the politically incorrect biases of those who create it.

 

[Cosmik]

...it would be very naive to think that AI won't also be affected by the politically incorrect biases of those who create it.

I would be hoping that once set going, the AI would create itself. But it will only happen if it is 'allowed' to make the actual associations that exist.

So, if an AI policeman cannot be allowed to appear biased against some designated victim group in society (it can always appear biased for them of course), then its innards will have to be tweaked by humans, or the data hand-adjusted by humans. I think it therefore falls over at the first hurdle, because if we are going for 'deep learning', humans cannot be expected to understand how it works. If humans must always be able to adjust it for acceptability of its results, we cannot have deep learning.

 

[maverickronin]

Is anyone aware of research into our brains' ability to compensate for equal loudness curves. It's well-established that the curves are basically correct and do not vary substantially from person to person (at least not in general trend), but given these curves are ever-present, it would seem surprising to me if we did not unconsciously compensate for them to a significant extent anyway.

IME, my brain doesn't doesn't seem to correct for it at all. I absolutely love the configurable loudness curve on my ADI-2 DAC since it makes low volume listening quite satisfying.

 

[andreasmaaan]

IME, my brain doesn't doesn't seem to correct for it at all. I absolutely love the configurable loudness curve on my ADI-2 DAC since it makes low volume listening quite satisfying.

Yeh, I have the same unit and have experimented with the same setting, but ultimately wasn't convinced it improved things. Are you listening through headphones or speakers with it?

 

[andreasmaaan]

I would be hoping that once set going, the AI would create itself. But it will only happen if it is 'allowed' to make the actual associations that exist.

As long as AI's goals are defined by humans, the biases of those humans will affect the AI. Moreover, I don't see necessarily that these biases won't continue to affect the AI after it begins to "create itself" either.

In any case, I don't like the idea of an AI cop whose goals are defined by AI.

And PS, I don’t always see it as a bad thing that human biases will be extended into AI, nor that some limits in terms of justice and fairnesswill (hopefully) be imposed by humans.

 

[maverickronin]

Yeh, I have the same unit and have experimented with the same setting, but ultimately wasn't convinced it improved things. Are you listening through headphones or speakers with it?

Just headphones, though from theory I'm not really sure why that would make a difference.

I didn't find it too hard to calibrate by ear either.

 

[andreasmaaan]

Just headphones, though from theory I'm not really sure why that would make a difference.

I didn't find it too hard to calibrate by ear either.

Interesting. I might have another go at it. Yeh, there's no reason from theory headphones vs. speakers should make a difference.

 

[maverickronin]

Interesting. I might have another go at it. Yeh, there's no reason from theory headphones vs. speakers should make a difference.

I set the bass and treble gain to 10 and started the Lo Vol Ref at what would be a quiet background listening level without the loudness engaged. Then just listened to some personal test tracks and adjust so that so that Lo Vol Ref and Lo Vol Ref + 10 don't seem overblown and Lo Vol Ref + 20 (where the boot drops back to zero) doesn't seem dull. Sometimes turn down the turn the treble gain down a few DB too.

I'm not going to claim it's perfect or anything though. The graphs in the manual show their curves don't match up with the real data as well as they could. Both filters are just shelf instead of parabolic (ish?) curves and the "treble" gain creeps way down into the midrange.

It's hardly 'reference' or something I'd recommend for mixing but I find it makes low volume listening much more enjoyable which is great for saving your ears.

 

[andreasmaaan]

I set the bass and treble gain to 10 and started the Lo Vol Ref at what would be a quiet background listening level without the loudness engaged. Then just listened to some personal test tracks and adjust so that so that Lo Vol Ref and Lo Vol Ref + 10 don't seem overblown and Lo Vol Ref + 20 (where the boot drops back to zero) doesn't seem dull. Sometimes turn down the turn the treble gain down a few DB too.

I'm not going to claim it's perfect or anything though. The graphs in the manual show their curves don't match up with the real data as well as they could. Both filters are just shelf instead of parabolic (ish?) curves and the "treble" gain creeps way down into the midrange.

It's hardly 'reference' or something I'd recommend for mixing but I find it makes low volume listening much more enjoyable which is great for saving your ears.

I tried a similar thing, with the treble boost toned down compared to the bass boost to try to mimic better the equal loudness curves. I think one reason I didn't persist with it much was that I rarely listen to headphones at low levels, i.e. headphone listening for me is usually at a pretty moderate SPL level. So I found that when the loudness function was active at a level I might actually listen at, it seemed like too much, while in the range where it was active, it always seemed to make things too bright.

I'm gonna revisit it though out of interest. Perhaps I should try without any treble boost at all...