Machine Learning for Analog Hardware Emulation

September 30, 2020


In a time where digital recording solutions are standard in most studios, a lot of people still prefer using analog hardware in some situations. If you ask them why you will most likely hear some terms like saturation, depth, or warmth or other fancy characteristics to justify the use of the old technology over modern ones. There is definitely something magical about analog recordings which is clearly not present the same way in the digital world. A few examples of such kind of hardware are guitar amps, microphone pre-amplifier or tape recorders. As one would expect, it has always been a goal for some people to combine the easy-to-handle nature of digital audio workstations with the analog sound of physical recording hardware. For instance, there are already tons of different software emulations for tube guitar amplifiers out there which try to achieve the same sound as traditional equipment.

Resonance EQ

Component-wise Modeling

The classic way to emulate hardware with software algorithms is to strip down the underlying circuits and model the properties of each component separately. This means the behavior of all parts like capacitors, inductors, transistors or tubes are analyzed and reconstructed in program code. These code modules can then be put together in the same way they are wired on a circuit board. Some of these approaches are actually getting pretty close to the real deal and clearly show a similar non-linear behavior as the real hardware. However, it is really difficult to model every little aspect and especially the frequency dependent interaction between components. Even if all the emulations of each component behave correctly, there are still many other factors like magnetic induction between wires or charging processes which affect the sound. These highly complex interactions can be partially described using differential equations and other sophisticated mathematical models. To improve the accurcy these calculations can be more and more refined which requires a lot of technical knowledge and gets really complicated and computionally expensive.
Resonance EQ

Machine Learning Based Modeling

The use of artificial intelligence techniques is on the rise in nearly all kind of industries and has shown to be able to find solutions to many until now unsolved problems. Nowadays, machine learning has become a standard in domains like image processing where very accurate classifiers and other tools could be built using artificial neural networks. It is an obvious idea to try to apply the same techniques to audio hardware emulation in order to overcome the mentioned limitations of human-made hardware models. In contrast to the component wise modeling, this technique sees the whole hardware as one system. This is useful since it takes every tiny interaction between elements into account and other details one might miss in a hand-crafted model. We explored this idea by building an artificial neural network which is capable of finding and learning the exact behavior of hardware circuits in all its nuances.
Resonance EQ

PreTube and PreFET

As a first proof of concept we dug out some analog pre-amplifiers and let the network listen to the input and output signals of the hardware to learn the exact mapping between both. This process is called training and requires tons of example data and computing resources. After some training iterations and a lot of modifications of the original neural network topology we were able to come up with some pretty cool sounding algorithms which clearly show the same characteristics as their hardware role-models. Here, the biggest limitation is the required computational power which we were able to reduce to a point where it can run on a single CPU even in multiple instances. Of course, using this very new approach there are still a lot of things to improve and refine but the first reactions we got so far definitely convinced us to further explore this technique:

„Sounding really great on drum loops and bass samples.“
„Getting a lot of use out of it in sound design.“
„The way it handled transients was quite convincing and dare I say analog.“
„A new era for analog modeling has begun!“
„This thing can do some of the extreme end of tube pre saturation quite convincingly.“
„It doesn't have that typical static plugin sound.“
„Very subtle effect but wonderful.“


PreFET is a free transistor pre-amp emulation while PreTube emulates three different tube pre-amps. Both plugins are available as VST3/AU/AAX and can be used to bring some analog magic to your tracks. It is also worth a try putting them on your master track to add some depth and warmth. We hope that in the future we can further improve this technique and come up with some more plugins of this kind.