Artificial Intelligence Filters in Streamers?

[Jungstar]

Can AI make budget sound like World Class?

Let's say I have a good budget system with DAC Sinad 120db, a decent 300w class D amp etc. Now, I'm looking to make it play like a dream $ 30k system with top 200 watt clean Class A/B amp and a fantastic 4K dac.

(ignore speakers for now)

If I played the same 100.000+ hours of music on both systems and we recorded it at ultimate settings, we would get a delta.
Ie what is the difference in output (ideally the amplified) on the Budget and the Dream System.

Now, using a filter, AI trained from the delta, I would try and get the budget system to play like the 30k system. I would apply the "inverse" filter to modify the electronic signal, so that the powered signal coming out of the affordable Amp was very close to the expensive system.

We already do this somehow when we upscale 1080p video to 4k. Basically, the AI was fed a million hours of video in 4k and 1080p and now learned to "fill in the missing parts" to get from 1080p to 4k. Here we would just like to taint the system to play like a world class DAC+Amp system.
So removing the unwanted and adding the wanted.

There must be some PhD students in AI and acoustics research....

 

[voodooless]

Why do you assume the 30k system is better than the 1k system? Let’s first establish that there is a audible difference at all. If so, all that is needed to make them match is some engineering.. no AI needed.

 

[JeffS7444]

To a limited extent, this is already being done routinely in the form of room correction where we try to adjust frequency response to match a target response. But this does not alter the speaker's polar response pattern to match that of another speaker, does not attempt to cancel speaker distortions, nor does it track your movements, save maybe for products like Apple's Airpods Pro which support it in specific situations.

But as far as sonic perfection in digital players, DACs and amplification is concerned, you can already achieve sonic transparency on a budget.

 

[Vincent Kars]

Reminds me a bit of the Bob Carver Challenge: https://www.stereophile.com/content/carver-challenge

IMHO audio don't need AI, audiophiles do.

 

[Jungstar]

Why do you assume the 30k system is better than the 1k system? Let’s first establish that there is a audible difference at all. If so, all that is needed to make them match is some engineering.. no AI needed.

Yes, I would assume the $30k system is better than a $2.5k system. But maybe it's more the speakers.. But as mentioned below, maybe it more of continuous room correction.

 

[sergeauckland]

Reminds me a bit of the Bob Carver Challenge: https://www.stereophile.com/content/carver-challenge

IMHO audio don't need AI, audiophiles do.

There's many an audiophile that could do with some human intelligence, let alone AI

S

 

[Katji]

The AI would be relatively free of psychology. ...iow, the logic problems.

 

[Katji]

"Artificial Intelligence Filters in Streamers?"

Filters? ...ok.
The topic/question refers to streamers - why streamers? - but then the story is about the possibility of using AI to make a complete system seem like an expensive/high-end system.

 

[ZgSTar]

Interesting... Acustica just released their new Sierra Rooms plugin. I suggest you guys try it. It works for headphones only because it's easy to know an almost exact frequency response. a similar approach could be used on speakers but you would need to measure the speaker in room response and start from there so not very plug and play but still interesting. You would not need to listen and sample the less performing rig but only sample the one you want to simulate. Remember however that if you don't put speakers in the equation it would never be possible even in a sci fi fiction scenario like the one you describe.
Above all else however give Sierra Rooms a try!

 

[Iglo]

To a limited extent, this is already being done routinely in the form of room correction where we try to adjust frequency response to match a target response. But this does not alter the speaker's polar response pattern to match that of another speaker, does not attempt to cancel speaker distortions, nor does it track your movements, save maybe for products like Apple's Airpods Pro which support it in specific situations.

But as far as sonic perfection in digital players, DACs and amplification is concerned, you can already achieve sonic transparency on a budget.

I agree with this point, I see little practical benefit of using "AI" just to make a system sound good, doing room correction or match a target frequency response. This is already possible without AI and for most users a set-and-forget affair.

There could be some benefit using "AI" (although I would specify machine learning in this case) for instance for large-scale concert sound systems. Tune the system to the desired response before the concert and have a ML-model continually correct for changes in the venue/field (more/less people-temperature/wind direction/...). I do think this would have a practical purpose and should be possible (if not already available).

The same technique could be used in a home situation to continuously tweak the system to your listening position (some sort of dynamic room correction). Obviously this has some practical challenges (how does the system know where the listener is in the room and at which height the ears are). All in all this adds a lot of complexity which offers little benefit in real life IMHO.

 

[PierreV]

What we really need is AI in cables. Quantum is so 20th century...

 

[somebodyelse]

Have a look at PedalNetRT and the papers and articles linked in its Info section for some examples of AI being used to reproduce distortion characteristics. It's not very efficient, and I'm not sure what the benefit would be over conventional convolution in the low distortion scenario here.

 

[usersky]

There is a trend to use AI for everythin on Earth, appropriate or not. I am sure that some people will burn some investor's money on such a project and I am also sure some audiofiles would find magic dimensions in the result. Reasonable people usualy do not use AI for 1+1 or FFT. But some will still try.

 

[Jungstar]

"Artificial Intelligence Filters in Streamers?"

Filters? ...ok.
The topic/question refers to streamers - why streamers? - but then the story is about the possibility of using AI to make a complete system seem like an expensive/high-end system.

it is not expensive at all. You can buy AI restoration programs for a powerful PC for 20 USD. OK the PC needs to be 1000 USD powerful.

 

[Jungstar]

Here is an AI image of what was Restored from the input.
That is also how your 4K TV works. Sending 4K real images over the internet is very taxing.
So we could also "upscale" sound like MP3 by running millions of hours of music through an AI or training a streamer or DAC to do continuous change to the sound. It would fill out the missing pieces correctly. And make it sound better or like a "better" stereo.
It is somewhat easier with a face, as faces tend to look similar. But it must and should be a huge focus on the next 10 years of Audio as the cost of processing AI is coming down like crazy. Like the current bitcoin boom is somewhat funding the AI as Graphics cards (who calculate AI) are coming down in price.

 

[Katji]

"Artificial Intelligence Filters in Streamers?"

Filters? ...ok.
The topic/question refers to streamers - why streamers? - but then the story is about the possibility of using AI to make a complete system seem like an expensive/high-end system.

it is not expensive at all. You can buy AI restoration programs for a powerful PC for 20 USD. OK the PC needs to be 1000 USD powerful.

smh. I did not say anything about the cost of AI. (Whatever.) Your topic is using AI to make a low-budget system seem like an expensive one.

And, nonsense. Bitcoin whatever using graphics processors is not "funding AI." The AI you're talkng about is software development - to put it simply. Nothing to do with cost of processing.

 

[PierreV]

Here is an AI image of what was Restored from the input.

Well, audio, even multi-channel audio, doesn't require a lot of bandwidth, at least on our side. Not sure it would be interesting to use AI upsampling/restoration on a lossless codec. As far as low quality MP3s are concerned, yes, maybe there would be an audible benefit, but which low quality MP3 is so unique that it wouldn't be possible to obtain the original lossless file?

I could see interesting stuff happening with historical or even bootleg concert recordings. Something like the AI-assisted coloring of old black and white movies.
But then, I am not sure I would like Hans Hotter to sound "modern".

 

[UKPI]

but which low quality MP3 is so unique that it wouldn't be possible to obtain the original lossless file?

Finding high quality soundtracks for old video games can be hard. Legally downloading those often isn't an option because either the right holders don't care enough or the music got caught in the "licensing purgatory". In this case, you have to buy a used disc, pirate them, or be content with the low quality audio files in the game.

Then again, not many areas would benefit from upscaling audio streams unlike upscaling videos. Newly encoded low bitrate audio files can sound okay without AI at this point. In AAC, technologies like SBR and Parametric Stereo do a reasonable job at suboptimal bitrates.

 

[Jungstar]

Maybe to upsample a Bluetooth stream.... but OK there, I'm looking for a problem to a solution.

Anyway, I was more interested in how to "color" sound to make an Entry System Sound Better.

This guy did a software filter to simulate a Tube Preamp as a bachelor thesis

https://philaudio.wordpress.com/research/tube-preamp/