Neural network processing genuinely has been a revolution in audio processing and analog emulation, the open source NAM and Proteus are amazing. Calling Machine Learning 'AI' still makes me cringe
@@blast_processing6577 Nope. I replaced all my Tape and Saturation to Proteus. It depends on the capture. Its like Convotution on Deeps learning basis. Next Level! The Baby Audio Tape and Hyper Technology from Acustica Audio is also based n that. You missed the revolution. I use The ceptures by PastToFutureReverbs.
@@pyratellamarecordingstudio1062 really depends on the quality of the model, and whether you run into it too hot. But yeah i use Two-Notes Genome, which is also neural modelling and supports NAM, AIDA-A, Proteus and its own models, and has about the best interface available. The ChowDSP ML software is amazing too
It's being marketed to the the last two generations that horrible parents produced...along with the horrible music produced...the horrible movies produced...the horrible etc. etc. etc...Imagine growing up and having B e i b e r as your childhood listening experience...;)
@@andothermoves well I did & I just chose not to listen... also you sound like one of the most horrible things ever produced, never seen that much negativity in one person...
Yeah as you said in 13:50.... the tape machine sounds so much better than the plugin. Especially the lows. And the hights too, because they are not so harsh. Great video and comparison!!
True. Kind of funny since you’d need to do more EQing on the tape machine results than on the plugin results post processing. The comparison expectation is kind idiotic in that sense.
The thing is I don’t think we know WHAT tape does. It’s frequency dependent, velocity, amplitude with harmonic character. All the devices I have tried are static.
The tape really warmed up the sound .. almost sounded like it opened up the music and brought it foward …. it made the plug-in almost sound a little narrow … Keep up the good work .. I like and appreciate your videos
That plugin really isn’t it 😂 It sounded so shrill compared to your tape machine. It does something nasty to the hi mids - almost sounds like a high Q EQ boost. Thanks for another great video.
to me it sounds like if the real thing had way more impact on lower frequencies adding the pleasant warmth and some like 3D feeling? while the vst resembled something like ozone sound giving it that digital signature feeling which isn't quite polished due to the harshness
The plugin seemed to lose the center image, or push the sides more, and comes across as harsher. The real tape keeps the mid channel up front and the top end stays smooth and pleasant.
im in my 50's, grew up listening to vinyl and tape. This sounds like a early consumer 1/4 inch 2 track reel from the 70's . which its perfect if you are trying to mimic that aged, old tech sound for period piece. i had a old akai reel and i used radio shack reels in it which sounded similar to that.
They didn't mention the tape machines they were training it on. Who knows, if the devices were already in a bad state to begin with and then AI just has to deal with SISO data (shit in, shit out).
just shows that the company has a serious case of Dunning-Krueger effect when it comes to this. otherwise, their plugins are quite good unfortunately. they've shot themselves on the foot with this one.
Only tape does tape, however, I just purchased the Victor-Custom 456 & it's staying on my mixbus! The craftsmanship is great & it sounds much better than plugins. His units are built for each customer & very affordable.
The term "AI" is enormously misused, for example there were server/PC cabinet manufacturers at Computex in Taiwan and AI was also widely advertised there. When they asked, what does that cupboard have to do with AI? They replied, yes, you can put an AI system in that cupboard. Blah lol. We also still have 2 Otari tape machines in the tracking studio. Much better than plugins. But indeed quite a lot of ssssssssssssssssssssssss that's what we filter out with Reaper's ReaFIR.
your tape machine sounded fuller to my ears. listening through a quality set of headphones on my laptop through youtube and you could still feel the difference in that the real machine wins.
Just to let you know, the tape M is completely broken when used in 48k. If that's what you're mixing in the results will sound terrible compared to what the plugins actually is meant to sound like.
Nowadays plugins companies just want to put the sticker Ai on their plugins as if it was some proof of quality. The problem is that using Ai becomes the first requirement and sounding good or sounding better than existing solutions is an afterthought. It’s like in the 90 where they would put lasers everywhere even when it doesn’t make sense like a widget that is supposed to help you stop smoking.
Never mind the plugin, they already hit snake oil territory when that promo video listed self driving cars as one of the things machine learning could already do. Jokes aside, my completely unverifiable gut feeling is that this is conventional tape emulation using EQ/waveshaping/etc. and the "AI" element is that they used machine learning to adjust the parameters of their algorithm to match a real tape machine until they felt it was close enough. So more of an "AI calibrated tape emulation." A totally legitimate, potentially really good use of machine learning - that's pretty much what the Sonnox plugins are doing, and I use a couple of them a lot (especially Smart:Limit). The difference there is that they're transparent about what's actually happening.
nothing pisses me off more than developers who build noise into their emulations without the ability to turn it off. it's all because of fetishization of analog equipment which is way more important to some people than a good sound. drives me nuts.
There is that classic soft but thumpy, large yet fitting low end of real tape - aswell as the overall smoothing of the audio, aswell as that dulled high end that needs to be boosted after. I'm not sure what was happening with the plugin, if you was using it at 48khz, there is a known bug
Apple uses a Neural Engine, correct. It’s the same thing as a Neural Network how it’s normally called by other chipsets from Intel, AMD, NVIDIA and so on. If we take a look at Mac computers with an M1 chip or higher, they can utilize the Neural Engine by using the Audio Unit (AU) plug-in format. The latest update for Logic Pro (Mac and iPad) is using AI for the Stem Splitter (really good stuff) and also the Chroma plugin and the virtual players. This is an example that developers can use a Neural Engine for plugins utilizing AI powered functions.
I really like the low mid width & punch that the real tape adds. It reminds me of the 495 hardware vs plugin comparison where the plugin also lacked the low mid width & punch that the hardware provided.
I use tape emulation plugins to simply add some "analog" color - not make my mixes sound as if they were recorded on a reel-to-reel. I find this to be a much reasonable, practical and honest way to look at "this whole thing." :)
Yeah I'd love to see this. Even if only to make myself feel better about having to redesign part of this dual LA2A I recently built from a clone curcuitboard and spending months calibrating, then fixing issues then recalibrating it over and over (I'm getting there, and it already sounded better than any plug emulation from first fire up, just I really want it to be spot on, so have sunk so much time into ironing out any tiny little quirk the clone circuit/wiring has).. :P
i wonder what the machine they modelled sounds like because i feel like this just didnt sound great in the first place without even comparing it to the real one
This manufacturer does actually talk about aliasing on their website and they say that, to deal with it, they use up sampling. But it’s not clear if the user simply needs to activate that (or even run the DAW at a higher sample rate…and then recalibrate the audio files) or if some automatic process upsamples to deal with it but they do mention it and imply that they are dealing with aliasing to ensure that their models stay matched to the hardware.
The tape machine brings a really nice punch to the snare and kick with a smoother overall top end, while the plugin has none of that punch in the snare and kick with a fizzy sounding top end.
Yeah the AI one is slightly brighter imo but it may be due to say the tapehead being cleaner or just the actual unit they used to test to is different by a slight margin as electronic components have +/- ratings on components usually like 1% 5% or 10% etc.. these values can make a difference in audio filter applications. Are these at the same gain levels or is the AI one slightly higher? You should snapshot eq to compare the outputs. Also I am wondering if the plugin runs at 32 bit float and what hz ratings it processes at? Analog is generally going to be as good as the ADA converter and its own SNR while the plugin in the box is going to be based upon the CPU or dsp engine running it.
My tape machine is only a tascam-tsr 8 (12 inch 8 track 7.5/15 ips) however it sounds amazing especially for dub and lo-fi. The problem is each tape is £100 (i wont use new/old stock, too much sticky shed on 456) and i need to spend money periodically maintaining the machine ( needs new roller tyres just now, and probably a capstan belt replacement) Also its a pain messing around degaussing it etc. mostly it just sits under a dust cover as it s a dog hair magnet !
I love my tape machines, but the noise can be left out as far as I'm concerned. The noise part is not something the actual model learned to do I think. Something about neural networks in audio processing. What's usually done is that you have an audio file you feed through the device you want to capture, this means there's a before and after. When training the network learns what to do to the 'before' to produce something similar to the target. Now we get to running the network on your plugin. We put in an audio sample and we get out a new audio sample. Just time domain processing and no FFT's. A network like this doesn't know anything but the data it's fed. This does fall within AI, and is usually referred to as Narrow Artificial Intelligence. This means it's good at a single task. The content of audio you feed to a network to train is greatly affecting the result, the type of network plays a large part as well. WAY more to talk about but this should uncover a little bit of the mystery Edit: not sure what they're doing with their APU's tbh
The real tape machine seems to be mellower. The AI sounds like it's saturating a lot more of the highs, but not in a good way. Almost just sounds like a light distortion plugin.
What about adding one additional stage of frequency curve adopting from the real tape. I quite like the sound of the plugin, but obviously the frequency response is totally different... I think the plugin would come much closer to the tape, when the frequency curve is the same...
Talking as a data scientist with master degree, their machine learning department really don’t know what they are talking about, it requires a lot of cpu or gpu (gpu mainly actually) to train your hundreds of parameters, even millions, based on training and test data. But when trained, these neural networks weights are actually not gonna retrain anymore, they know their weight and won’t readjust afterwards. I actually think tape emulation with neural network is a thing and can work, but the issue is not the model itself, but what parameter you allowed to the model and if you have too much of them, then it can be RAM and Cpu intensive, this is why you then penalize your model to have less freedom and still behave as a tape
On emulation plug-ins, I often don’t choose “Noise ON” (because I just don’t like THAT noise) but I have a hard time agreeing with automatic “NO NOISE” approach. Because it’s part of analog…& we hate how digital never sounds like analog, while we laugh at ourselves for previously trying to make real-analog sound less noisy. Well let’s get rid of dither 🤷♂️ My point is the analog noise was doing something pleasing, even if we were gating it, etc.
painful highs even through telephone speakers... I'm very sensitive to these frequencies tho, but man, the sound changes too much and the vibe as well.
I could hear the difference between the two even over UA-cam on my iPhone speakers with my eyes closed. But then again, every tape machine Sounds different, what Unit did they use to train the ai? What tape did they use? So many variables…
Almost sounds like bitcrushing. I feel the plugin developers went as close as they could without thinking about aliasing and put the noise there as a quick fix instead of fixing the root cause problem. Honestly this shows that using AI still means you have to understand signal processing on a high level to implement things. These developers did not. I guess they basically have to redo the whole tech considering any regard to aliasing is missing, so this is almost like a way to get something out of a failed experiment instead of an actual product.
The JH-24 sounds better than the plugin, but the original (bypassed) sounds better than either. It seems to me that they are 'Turning-Off' more than just the noise alone, when you turn off the 'Noise', they can fool people into thinking that it's the Noise making it sound better, when it's really just anti-aliasing being turned ON.
Yow Man ive been struggling to hear the difference when using tape machine plugin can you make a video that showing or hearing the real thing what is happening when using a tape machine
So when using any sort of tape plugin, should you not insert one at the bottom of your FX rack (recording to tape, assuming all analog plugins/pedals/boxes) and ALSO on your master track (printing of master tape to the stereo tape)?
The free Cubase tape machine vst sounds better than that. I understand the power of NN for complex dynamic variable problems like effective noise reduction, but for a deterministic process like tape saturation? I think is better and CPU-friendlier to just model the proper transfer function and add some noise.
I mean, this is entirely a matter of taste, but I’ve never been sure why everyone’s so nuts about tape saturation in the first place; valve saturation has always sounded so much sweeter to me because of all the even-order harmonics.
Since the eq of the real tape and the plug-in are different, and the real tape had a nicer darker eq for the song, I dunno if the a b comparison was very useful.
Analog will always contain real soul, inherit to natural imperfections and real non aliased waveform curvature. Got a reel to reel a few months ago and never going back for final prints.
In my experience, 0’s and 1’s can’t compete with sending audio through actual circuits. It all adds up when tracking all parts through analog and then printing the final mix. If you record everything digital and then AB the finished product, the difference is not subtle to my ears.
The plugin’s hi-mids are like sandpaper in my headphones, your tape was butter. If the devs did any AB testing, putting this out for people to purchase is deceptive. With that output and the weird promo video, I get the vibe that this one might go beyond the typical overhyped startup snake oil into a full blown scam of a development. I’d put money that there isn’t even any machine learning involved 😅😅😅
I should have thought that tech based on impulse responses and extrapolations between them would have been more promising for catching the qualities of tape in digital. I know that Acustica Audio and I think Overloud have taken that sort of approach. There would still be aliasing, so some anti-aliasing approach would still be necessary. Perhaps machine learning could have some application in developing impulse response-based technology. But this approach of merely trying to reproduce the output with different inputs over many hours might be a little bit of a blunt use of an otherwise sophisticated approach.
MCI sound very well with solid low end and smooth on top, but I think it makes no sense to compare it. The vst is not based on your tape machine with your tape and your calibration.
Here’s the thing about neural networks. They’re capable of expressing some really gnarly functions. And they’re hard to train well. It’s hard for a lot of inexperienced ML scientists to even know whether they’ve trained them well. These audio processing modules (APU’s) - i don’t know they even really know how to train these well. The -70db error - i want to know more. I wonder about what data they evaluated it on. People mess this up all the time. Is this on the training set? the teat set? You never know with ML scientists because mediocre ones mess this stuff up all the time. I think it’s possible to do this well, but it’s difficult. I’d be concerned that it would do weird, unsmooth things to my audio; feed in simple waveforms and get out very weird, subtle deformations / distortion that weren’t trained or regularized out. If it’s just a bunch of APU’s - then it sounds like rather than leveraging the amazing power of real NN’s it’s more like approximating NN’s via a smaller number of potentially crappy but probably basic plugins over and over in series. Interesting ideas, skeptical about execution.
Have you seen Synplant synth when using the AI copy mode? It's really really well done and any straight oscilator synth sound it can emulate almost perfectly, automatically (even ones with more oscilators than it it's self actually has -up until a point ofc). Any complex multi filter modulation throws it off though, but it's not really designed for that, but future plugs using this tech from the same guy may be more indepth and be able to recreate even more complex sounds. The designer did a great vid about it's creation, and to me it's the best implamentation of machine learning in a plugin so far.
Hi! It would you mind running the training signal through your tape machine so i can run train Neural amp modeler with it? There is just one Studer available now and it sounds good but it was done in a rush i think. Let me know if youd like to set it up, it would be really cool!!!
@@sparella Ah! smart. But i've used it and didnt get the warmth that im hearing trough his tape machine. Also, profiling other vsts etc is concidered wrongdoing so maybe thats why
Neural network processing genuinely has been a revolution in audio processing and analog emulation, the open source NAM and Proteus are amazing. Calling Machine Learning 'AI' still makes me cringe
Open source NAM and Proteus? That isn't an emulation of the Proteus 2000 by any chance, is it?
@@blast_processing6577 Nope. I replaced all my Tape and Saturation to Proteus. It depends on the capture. Its like Convotution on Deeps learning basis. Next Level! The Baby Audio Tape and Hyper Technology from Acustica Audio is also based n that. You missed the revolution. I use The ceptures by PastToFutureReverbs.
NAM is awesome, but still has alias issues. Proteus is ok.
@@pyratellamarecordingstudio1062 really depends on the quality of the model, and whether you run into it too hot. But yeah i use Two-Notes Genome, which is also neural modelling and supports NAM, AIDA-A, Proteus and its own models, and has about the best interface available.
The ChowDSP ML software is amazing too
I hear you on the cringe. I fear we have reached the point of no return. 🤕
did they also replace their marketing department with an ai? WTF was that 2012 Machinima music with TTS voice over shitty annimations video lol...
Better than the other marketing videos that all look the same.
@@Cynshet yeah the classic "being different by throwing the marketing budget in the trash" approach
The minute I heard "symphony" alarm bells went off
It's being marketed to the the last two generations that horrible parents produced...along with the horrible music produced...the horrible movies produced...the horrible etc. etc. etc...Imagine growing up and having B e i b e r as your childhood listening experience...;)
@@andothermoves well I did & I just chose not to listen... also you sound like one of the most horrible things ever produced, never seen that much negativity in one person...
Good to see people still making high quality porn soundtracks.
Yeah as you said in 13:50.... the tape machine sounds so much better than the plugin. Especially the lows. And the hights too, because they are not so harsh. Great video and comparison!!
True. Kind of funny since you’d need to do more EQing on the tape machine results than on the plugin results post processing. The comparison expectation is kind idiotic in that sense.
The thing is I don’t think we know WHAT tape does. It’s frequency dependent, velocity, amplitude with harmonic character. All the devices I have tried are static.
The tape really warmed up the sound .. almost sounded like it opened up the music and brought it foward …. it made the plug-in almost sound a little narrow …
Keep up the good work .. I like and appreciate your videos
That tambourine sounded harsh AF. Keep up the good work!
"a lot of aliasing" - aliasing doesn't even start above -110 dbs.
Its fine in terms of aliasing.
That plugin really isn’t it 😂 It sounded so shrill compared to your tape machine. It does something nasty to the hi mids - almost sounds like a high Q EQ boost. Thanks for another great video.
Satin obliterated that thing. You can't beat proper modelling.
to me it sounds like if the real thing had way more impact on lower frequencies adding the pleasant warmth and some like 3D feeling? while the vst resembled something like ozone sound giving it that digital signature feeling which isn't quite polished due to the harshness
The plugin seemed to lose the center image, or push the sides more, and comes across as harsher. The real tape keeps the mid channel up front and the top end stays smooth and pleasant.
Agreed.
You're absolutely deaf.
I'mma keep using MixAnalog until the wheels fall off their server racks!
im in my 50's, grew up listening to vinyl and tape. This sounds like a early consumer 1/4 inch 2 track reel from the 70's . which its perfect if you are trying to mimic that aged, old tech sound for period piece. i had a old akai reel and i used radio shack reels in it which sounded similar to that.
They didn't mention the tape machines they were training it on. Who knows, if the devices were already in a bad state to begin with and then AI just has to deal with SISO data (shit in, shit out).
just shows that the company has a serious case of Dunning-Krueger effect when it comes to this. otherwise, their plugins are quite good unfortunately.
they've shot themselves on the foot with this one.
Only tape does tape, however, I just purchased the Victor-Custom 456 & it's staying on my mixbus! The craftsmanship is great & it sounds much better than plugins. His units are built for each customer & very affordable.
His face in the thumbail and the title declaring that "this isnt it" is hilarious. 😂
The term "AI" is enormously misused, for example there were server/PC cabinet manufacturers at Computex in Taiwan and AI was also widely advertised there. When they asked, what does that cupboard have to do with AI? They replied, yes, you can put an AI system in that cupboard. Blah lol. We also still have 2 Otari tape machines in the tracking studio. Much better than plugins. But indeed quite a lot of ssssssssssssssssssssssss that's what we filter out with Reaper's ReaFIR.
your tape machine sounded fuller to my ears. listening through a quality set of headphones on my laptop through youtube and you could still feel the difference in that the real machine wins.
Just to let you know, the tape M is completely broken when used in 48k. If that's what you're mixing in the results will sound terrible compared to what the plugins actually is meant to sound like.
What sample rate should it be used on and why would it underperform so badly at 48khz?
@@EpithetMusicTV it's a widely reported bug which they are trying to sort out. Not sure it's like that on all systems.
If that is what is happening here then he needs to replace the video.
@@millenniummastering that's why I thought I'd mention it 😎 It's a great plugin at 44,1khz, one of my favourites atm.
No one uses 44.1 anymore. If your plugin can't run at 48, you shouldn't release it.
Dude, as soon as you put on the "real" tape machine , the kick become fat, nice rounded and puffy 😌amazing...
Yes, please do make the alignment video. Would come in very handy, as I have the exact same machine here :) I'm planning on recapping the PSU's soon.
Nowadays plugins companies just want to put the sticker Ai on their plugins as if it was some proof of quality. The problem is that using Ai becomes the first requirement and sounding good or sounding better than existing solutions is an afterthought. It’s like in the 90 where they would put lasers everywhere even when it doesn’t make sense like a widget that is supposed to help you stop smoking.
Now they have a real reference with your actual machine. I think they will be very thankful for your test and be able to fix! 👍😉
Now they have a real reference with your actual machine.
I think they will be very thankful for your test and be able to fix! 👍
Just sounds like crunch
Never mind the plugin, they already hit snake oil territory when that promo video listed self driving cars as one of the things machine learning could already do.
Jokes aside, my completely unverifiable gut feeling is that this is conventional tape emulation using EQ/waveshaping/etc. and the "AI" element is that they used machine learning to adjust the parameters of their algorithm to match a real tape machine until they felt it was close enough. So more of an "AI calibrated tape emulation."
A totally legitimate, potentially really good use of machine learning - that's pretty much what the Sonnox plugins are doing, and I use a couple of them a lot (especially Smart:Limit). The difference there is that they're transparent about what's actually happening.
they could have just gone the NAM route and try to do some original research
nothing pisses me off more than developers who build noise into their emulations without the ability to turn it off. it's all because of fetishization of analog equipment which is way more important to some people than a good sound. drives me nuts.
There is that classic soft but thumpy, large yet fitting low end of real tape - aswell as the overall smoothing of the audio, aswell as that dulled high end that needs to be boosted after.
I'm not sure what was happening with the plugin, if you was using it at 48khz, there is a known bug
for some reason the real tape machine seems to be making the stereo image wider as well
Apple uses a Neural Engine, correct. It’s the same thing as a Neural Network how it’s normally called by other chipsets from Intel, AMD, NVIDIA and so on. If we take a look at Mac computers with an M1 chip or higher, they can utilize the Neural Engine by using the Audio Unit (AU) plug-in format. The latest update for Logic Pro (Mac and iPad) is using AI for the Stem Splitter (really good stuff) and also the Chroma plugin and the virtual players. This is an example that developers can use a Neural Engine for plugins utilizing AI powered functions.
Isn't that Sandworm by Ja Rule?
Your tape machine has much more presence in the lows and mids. It sounds way better.
Yes the plugin sounds bad... your tape machine is king
PeaceR " )
Why more lows and mids sound better?
@@TheHmmka To my taste, it adds what's missing in the dry signal.
@@DrKaoliN did he oversample it?
@DrKaoliN - absolutely: a much fuller sound.
I really like the low mid width & punch that the real tape adds. It reminds me of the 495 hardware vs plugin comparison where the plugin also lacked the low mid width & punch that the hardware provided.
I use tape emulation plugins to simply add some "analog" color - not make my mixes sound as if they were recorded on a reel-to-reel. I find this to be a much reasonable, practical and honest way to look at "this whole thing." :)
I always hear how snare drums are really alive on tape vs the digital whatever version. Tape is unbeatable
Interesting. You can definitely hear the real tape rolling off some high end. The VST sounds brighter.
Would love to see a Tape Machine alignment video
yea make a video of you calibrating the tape machine dope video
Yeah I'd love to see this. Even if only to make myself feel better about having to redesign part of this dual LA2A I recently built from a clone curcuitboard and spending months calibrating, then fixing issues then recalibrating it over and over (I'm getting there, and it already sounded better than any plug emulation from first fire up, just I really want it to be spot on, so have sunk so much time into ironing out any tiny little quirk the clone circuit/wiring has).. :P
The noise is part of the analog sound.
i wonder what the machine they modelled sounds like because i feel like this just didnt sound great in the first place without even comparing it to the real one
There is a known bug if using 48k. Used it at 44.1k without these issues. The developers are looking to sort the bug. Definitely a plugin to watch.
no not really, much better options
This manufacturer does actually talk about aliasing on their website and they say that, to deal with it, they use up sampling. But it’s not clear if the user simply needs to activate that (or even run the DAW at a higher sample rate…and then recalibrate the audio files) or if some automatic process upsamples to deal with it but they do mention it and imply that they are dealing with aliasing to ensure that their models stay matched to the hardware.
They're good in both ways, the real tape sounds darker and fuller in lows while still gives that tape power
cana san martin plugins?
The tape machine brings a really nice punch to the snare and kick with a smoother overall top end, while the plugin has none of that punch in the snare and kick with a fizzy sounding top end.
yes it comes accross youtube, can hear the difference.
I in fact do think it'd be cool to see a video about calibration/alignment of a tape machine. I never got to work with them, so I'm always intrigued.
an imitation will never be as good as the real thing
I’ll hang on to my neve 542’s a little longer…
Yeah the AI one is slightly brighter imo but it may be due to say the tapehead being cleaner or just the actual unit they used to test to is different by a slight margin as electronic components have +/- ratings on components usually like 1% 5% or 10% etc.. these values can make a difference in audio filter applications. Are these at the same gain levels or is the AI one slightly higher? You should snapshot eq to compare the outputs. Also I am wondering if the plugin runs at 32 bit float and what hz ratings it processes at? Analog is generally going to be as good as the ADA converter and its own SNR while the plugin in the box is going to be based upon the CPU or dsp engine running it.
seems like they put in a smile shaped EQ curve as well.
Ahhaaaa I know EXACTLY what the problem was with the plugin. How long did you leave it to warm up first :))
It's exactly like tape, only with slightly bit-crushed, aliased highs.... And slightly crunchy and smeared lows (crunchy transients in general)
My tape machine is only a tascam-tsr 8 (12 inch 8 track 7.5/15 ips) however it sounds amazing especially for dub and lo-fi. The problem is each tape is £100 (i wont use new/old stock, too much sticky shed on 456) and i need to spend money periodically maintaining the machine ( needs new roller tyres just now, and probably a capstan belt replacement)
Also its a pain messing around degaussing it etc. mostly it just sits under a dust cover as it s a dog hair magnet !
I love my tape machines, but the noise can be left out as far as I'm concerned. The noise part is not something the actual model learned to do I think.
Something about neural networks in audio processing. What's usually done is that you have an audio file you feed through the device you want to capture, this means there's a before and after. When training the network learns what to do to the 'before' to produce something similar to the target. Now we get to running the network on your plugin. We put in an audio sample and we get out a new audio sample. Just time domain processing and no FFT's. A network like this doesn't know anything but the data it's fed. This does fall within AI, and is usually referred to as Narrow Artificial Intelligence. This means it's good at a single task. The content of audio you feed to a network to train is greatly affecting the result, the type of network plays a large part as well. WAY more to talk about but this should uncover a little bit of the mystery
Edit: not sure what they're doing with their APU's tbh
The real tape machine seems to be mellower. The AI sounds like it's saturating a lot more of the highs, but not in a good way. Almost just sounds like a light distortion plugin.
What about adding one additional stage of frequency curve adopting from the real tape. I quite like the sound of the plugin, but obviously the frequency response is totally different... I think the plugin would come much closer to the tape, when the frequency curve is the same...
Compared to UAD Ampex ATR-102 and it is really close, if you compensate with EQ... so conclusion for me: Don't have to buy it :-)
They probably used a Tascam tape machine running at 7.5ips😮
The plugin was so freaking harsh while also removing depth and punch. ^^
You going to make a video on kraftur by soundtheory soon? (The gullfoss makers)
SSL E series with noise gate on every channel. Whoot Whoot!
Talking as a data scientist with master degree, their machine learning department really don’t know what they are talking about, it requires a lot of cpu or gpu (gpu mainly actually) to train your hundreds of parameters, even millions, based on training and test data. But when trained, these neural networks weights are actually not gonna retrain anymore, they know their weight and won’t readjust afterwards.
I actually think tape emulation with neural network is a thing and can work, but the issue is not the model itself, but what parameter you allowed to the model and if you have too much of them, then it can be RAM and Cpu intensive, this is why you then penalize your model to have less freedom and still behave as a tape
youtube codec only lowpasses at 17 KHz in certain circumstances but in normal circumstances you get full 256 kbps opus audio low passed only at 20K
On emulation plug-ins, I often don’t choose “Noise ON” (because I just don’t like THAT noise) but I have a hard time agreeing with automatic “NO NOISE” approach. Because it’s part of analog…& we hate how digital never sounds like analog, while we laugh at ourselves for previously trying to make real-analog sound less noisy. Well let’s get rid of dither 🤷♂️
My point is the analog noise was doing something pleasing, even if we were gating it, etc.
Check out airwindows noise shapers
Man, that smooth, lovely sound of the Sony 😍
Please do the tape realignment vid pls
When I trialed it I couldn’t stand how the GUI worked. I don’t think there was a way to manually type in values which drove me nuts.
painful highs even through telephone speakers... I'm very sensitive to these frequencies tho, but man, the sound changes too much and the vibe as well.
Every saturator is harsh on high frequencies if you dont oversample it by at least 8x. Just load the saturator on superplugin vst and oversample it.
It seems that with normal tape machine the sound effects ower all and the plug in only effect the high sound
WOULD LOVE TO SEE YOU REVIEW THE GOLD CLIP PLUGIN
I could hear the difference between the two even over UA-cam on my iPhone speakers with my eyes closed. But then again, every tape machine Sounds different, what Unit did they use to train the ai? What tape did they use? So many variables…
Almost sounds like bitcrushing. I feel the plugin developers went as close as they could without thinking about aliasing and put the noise there as a quick fix instead of fixing the root cause problem. Honestly this shows that using AI still means you have to understand signal processing on a high level to implement things. These developers did not. I guess they basically have to redo the whole tech considering any regard to aliasing is missing, so this is almost like a way to get something out of a failed experiment instead of an actual product.
The JH-24 sounds better than the plugin, but the original (bypassed) sounds better than either. It seems to me that they are 'Turning-Off' more than just the noise alone, when you turn off the 'Noise', they can fool people into thinking that it's the Noise making it sound better, when it's really just anti-aliasing being turned ON.
so if you leave the noise on and the denoise the plugin it would sound more realistic right?
where is the bottom end gone on the plugin, for my it sound like a cassette with chrome tape in it
"might as well start flames over how arse the plugin is and stop it from spreading." - admin シンジョ
"edit: hard power is an option here, thus."
Yow Man ive been struggling to hear the difference when using tape machine plugin can you make a video that showing or hearing the real thing what is happening when using a tape machine
So when using any sort of tape plugin, should you not insert one at the bottom of your FX rack (recording to tape, assuming all analog plugins/pedals/boxes) and ALSO on your master track (printing of master tape to the stereo tape)?
theres a stray peak on the tambo poking out thats getting driven. needs a little fabfilter notch
You can hear the compression on the real tape and there's nothing on the plgun. I use Slates VTM most of the time and I think it sounds pretty good. 😎
The real tape has an emphasis on the low mid and have a bit more presence and image..
any opinions on the UAD Ampex? its probably the best tape emulation i heard so far
I thought the Low Pass Filter thing only occurs, when you rip something off YT. But in the App the Audio is fine. 🤔
The free Cubase tape machine vst sounds better than that. I understand the power of NN for complex dynamic variable problems like effective noise reduction, but for a deterministic process like tape saturation? I think is better and CPU-friendlier to just model the proper transfer function and add some noise.
What kind of tape do you use on your tape machine?
I mean, this is entirely a matter of taste, but I’ve never been sure why everyone’s so nuts about tape saturation in the first place; valve saturation has always sounded so much sweeter to me because of all the even-order harmonics.
Since the eq of the real tape and the plug-in are different, and the real tape had a nicer darker eq for the song, I dunno if the a b comparison was very useful.
Analog will always contain real soul, inherit to natural imperfections and real non aliased waveform curvature. Got a reel to reel a few months ago and never going back for final prints.
The original sounds warm and lively. The plugin is cold and dead and distorts the highs.
Thanks for the Video!
I would still like you to try the UA studer plugin and the Kalifornia Dynamic Alkaline Saturation plugin. Keep pushing 🎉
Not sure if you have Logic but have you tried out Chromaglow yet?
Your tape machine just sounds like it has that EQ bump in the low end.
ya.. for whatever reason the low-end from the tape machine was COMPLETELY different... and more pronounced.
that is what tape does...
@@morbidmanmusic Tape can be adjusted and calibrated.
that's all? you did not notice the warmer upper frequencies of reel tape, the harness of the plugin
Real tape 4life, yo
the sony tape has more weight and natural stereo width, the mid high rage is different.. and the vst sound more bright..
In my experience, 0’s and 1’s can’t compete with sending audio through actual circuits. It all adds up when tracking all parts through analog and then printing the final mix. If you record everything digital and then AB the finished product, the difference is not subtle to my ears.
The plugin’s hi-mids are like sandpaper in my headphones, your tape was butter. If the devs did any AB testing, putting this out for people to purchase is deceptive. With that output and the weird promo video, I get the vibe that this one might go beyond the typical overhyped startup snake oil into a full blown scam of a development. I’d put money that there isn’t even any machine learning involved 😅😅😅
I should have thought that tech based on impulse responses and extrapolations between them would have been more promising for catching the qualities of tape in digital. I know that Acustica Audio and I think Overloud have taken that sort of approach. There would still be aliasing, so some anti-aliasing approach would still be necessary. Perhaps machine learning could have some application in developing impulse response-based technology. But this approach of merely trying to reproduce the output with different inputs over many hours might be a little bit of a blunt use of an otherwise sophisticated approach.
Having a better tape machine is such a oxymoron if i ever heard one 😂
MCI sound very well with solid low end and smooth on top, but I think it makes no sense to compare it.
The vst is not based on your tape machine with your tape and your calibration.
whats the best tape machine emulator right now? I heard the t-racks stuff is good, also the UHE tape stuff, any other recommendattions?
A lot of people like Slate VTM. The UAD stuff is well respected. For a more noticeable effect, Taupe is pretty good. Yes, the IK stuff is very good.
Here’s the thing about neural networks. They’re capable of expressing some really gnarly functions. And they’re hard to train well. It’s hard for a lot of inexperienced ML scientists to even know whether they’ve trained them well.
These audio processing modules (APU’s) - i don’t know they even really know how to train these well.
The -70db error - i want to know more. I wonder about what data they evaluated it on. People mess this up all the time. Is this on the training set? the teat set? You never know with ML scientists because mediocre ones mess this stuff up all the time.
I think it’s possible to do this well, but it’s difficult.
I’d be concerned that it would do weird, unsmooth things to my audio; feed in simple waveforms and get out very weird, subtle deformations / distortion that weren’t trained or regularized out.
If it’s just a bunch of APU’s - then it sounds like rather than leveraging the amazing power of real NN’s it’s more like approximating NN’s via a smaller number of potentially crappy but probably basic plugins over and over in series.
Interesting ideas, skeptical about execution.
Have you seen Synplant synth when using the AI copy mode? It's really really well done and any straight oscilator synth sound it can emulate almost perfectly, automatically (even ones with more oscilators than it it's self actually has -up until a point ofc). Any complex multi filter modulation throws it off though, but it's not really designed for that, but future plugs using this tech from the same guy may be more indepth and be able to recreate even more complex sounds. The designer did a great vid about it's creation, and to me it's the best implamentation of machine learning in a plugin so far.
wanna see calibration video
Hi! It would you mind running the training signal through your tape machine so i can run train Neural amp modeler with it? There is just one Studer available now and it sounds good but it was done in a rush i think. Let me know if youd like to set it up, it would be really cool!!!
I'm surprised no one has profiled Mix:Analog's tape machines yet.
@@sparella Ah! smart. But i've used it and didnt get the warmth that im hearing trough his tape machine. Also, profiling other vsts etc is concidered wrongdoing so maybe thats why