Hi, I'm very interested in the machine learning method of capturing and emulating analog devices. Recently a company called Analogxai posted a christmas sale on these fora, this sparked my interest in their process.

Looks like the hardware is sampled from here:
https://mixanalog.com/

Using this tech :
https://guitarml.com/
https://guitarml.com/#products

Tutorial for creating captures/models:

I think the only missing piece here is the set of data to use to send through the analog device, sample the result and present that to the machine learning algo. In the pedal example in the above video, it's trained using a set of guitar recordings (included with the utility provided by GuitarML/Proteus).
So I assume if we wanted to do the same thing but for a broader use case (ie; any instrument) we'd need a data set more akin to the kind of input that Acustica or other IR based products must have used to capture the hardware.

This would be pretty easy to test if anyone here has knowledge of a good set of recordings to use as the 'impulse'.
Mixanalog even have a free trial to send it through their analog gear.

Because most of these steps are facilitated by freeware, I think it's probably fine to talk about how we as a community might make our own models and share.

 

Last edited: Dec 13, 2023

Hi, thanks for your interest in AnalogXAi.

I do utilise LSTM deep learning network, however the code I use is customised for capturing analog hardware and I believe (any over 1000 customers so far) this is what sets our stuff apart from other stuff out there. Critical phase alignment amongst other experience I have picked up through 100's of hours of sampling high end hardware. The training file created by Keith Bloom is perfectly fine although you can devise your own training files. The open source code is open to anyone that wants to invest their time and energy into this tech from either a free or commercial platform.

The profiles I produce are dynamic and react as it would if you drive the signal harder into the hardware device, so saturation increases etc. This is due to the way that the sampling process and black box type modelling works, it is amazing and I spend the majority of my time doing this now, it is very time consuming to produce end to end lots of preset variations, testing / A/B and so on so that users can enjoy a wide range of high end analog processing options to enhance their productions.

We have sampled hardware products via cloud means I have also spent a lot of money and time with them too ;-) however I have also sampled and own my own high end hardware and also have sampled other friends hardware too. I have a big project starting in the new year with a major music company in London, think DW Fearn gear, consoles etc ;-)

I use BYOD because it has wet dry mix and gives me zero PDC latency, yes i can track through all the profiles without delay, unlike NAM it is stereo off the bat and I can run 100's of these plugins across big projects and still be only utilising 20% across my i9 12th Gen PC. I also advise to use reaper due to its ability to utilise all your CPU cores. next level mixing and processing power on tap ;-)

I hope this helps.

Best Wishes
Stuart

AnalogXai

 

Hi Stuart, how does your tech compare to Acustica's? I know yours is more of a snapshot, but is it still similar in to how the gear is captured? In a non biased way (tough i know) how does yours differ? what are some pro's of analogxai vs acustica, what are some cons? Do you have plans to create more "tweakable" programs? I tried the demo suite out of yours, and it sounds great, im going to grab a couple packs on the sale.

 

Hi Shinyzen,

Firstly thanks for takingthe time out to try my profiles.

Acustica Pros: Shiny photo realistic GUI's, extremely tweakable, use of many samples within one plugin, many sampled hardware devices to chose from, some sound pretty good actually,

Acustica Cons: Protection - I wont go into this too much but its well documented across many sites over the last few weeks and their tech is improving through use of algo which is great for reducing CPU usage but a lot of the plugins have issues and appear to take up massive amounts of HD space due to current protection. Compression doesn't sound great or analog. latency, not great and tracking with them using ZL version of their plugins takes significant CPU for a lot of users. Not the cheapest plugins on the market

AnalogXAi Pro's: Ask our users ;-) but I say we sound closer to the actual device we sampled because we use newer customised Deep learning technology (opensource) that has turned many heads in the world of Guitar Amplifier and Pedal cloning.
CPU - we use BYOD which is a free plugin host with 0PDC latency reported in your DAW host, so you can track with the profiles and use 100's in projects without CPU strain. Cheap, not expensive to buy aand lots of analog hardware not available in plugin format
Many presets to cover a wide range of common studio applications.
0 protection, we rely on the faith of our customers.
Dynamic black box sampling - you can drive the profile shard and they do not break up like most digital plugins...

AnalogXAi Cons: No flashy GUI but this reduces CPU and helps to achieve a 0PDC latency.
Simplified workflow, drive the profile to the sweet spot, till it sounds great and move on ;-)
Newer company so less well known
Oversampling required if using other than 44.1 in your DAW

I hope that helps

 

Last edited: Dec 13, 2023

thank you for the insight! still deciding on what to grab, but im going to snag some later today.

 

Maybe one major drawback of the ML / Proteus workflow is that you are essentially working with preset snapshots.

I never use presets when working with EQ. It's always start flat and tweak according to the sound of the audio. This can be achieved with Acustica tech but not with ML / Proteus right now. At least not in a efficient way (you could, arguably, use many ML instances for each band but this is obviously hugely unwieldy).

So currently I love ML for tape, saturation, colour. Not so much for things that need to be tweaked to find the sweet spot for the particular audio you are currently working on. So compression, EQ.. Things like that.

 

@StUtOpi4 how do the analogxai captures work as far as sample rates? are they limited to one sample rate?

 

if using other than 44.1 just oversampling using industry standard oversampling programs ;-)

 

Very poor workflow for a very subtle gain that can be easily achieved with other methods. For example, combining TrueIron (Karzog), a Soothe, and opening up the stereo with StageOne. There you go, a plugin from Analog Obsession or UAD will sound analog. This combo I mentioned on Acustica Audio is deadly; hardly anyone will be able to determine in an A/B test against hardware which sounds more analog.

AnalogXAI is an interesting toy but impractical for everyday use. The interface still needs a lot of improvement

 

Interesting, I'd like to read about that. Where should I look?

 

Of course everyone is entitled to their own opinion but I focus bringing true analog sound to the DAW table and the majority of the AnalogXAi users and customers think i'm doing a very good job of it. I'd argue that the simple workflow is liberating in that it is as simple as using an 1176 fixed threshold compressor... just drive into it till you're happy, adjust the output and done... The profiles I do are adding saturation (typically the stuff that doesn't break up like shitty digital saturation), Compression, tape, mic pre's and some EQ.

Dont get me started on flashy GUI's that help you "think" your getting a good tone and simply add bloat to plugins, slow then down and cause no end of bugs... Since using AnalogXAi profiles I actually started to use my ears and stopped mixing by sight... Its an interesting concept and a lost art if you ask me ;-)

Happy Mixing and every man to their own.
Stuart

AnalogXAi

 

so i bought a couple of the packs, and am disappointed with the Oversampling. I tried DDMF Metaplugin, Superplugin, Blue Cat Patchwork, and Reaper. ALL of them displayed frequency loss when oversampling. I made sure the correct settings in BYOD were engaged, no oversampling, etc, but yah, oversampling does not work. The snapshots do sound good when not oversampled, however, not useable for me really, as i dont work at 44.1 Are there any plans for higher sample rates in the future?