We’re going to talk about product design, a subject I’m currently very passionate about, and one that some developers occasionally discuss, such as Tomislav Zlatic from Bedroom Producers Blog + Flame Sound regarding the release of his plug-in Vatra.

You are about to read about the process I used to select the plug-in’s final controls, in which little to no AI is used (I am still considering what I want to do with these technologies and what would constitute “fair use”, which will be the subject of another blog post), how I laid out the graphical interface, the elements I focused on for this new competition entry like accessibility and sweet spots, the reasons behind this unconventional concept (reverb + instrument), and my general workflow when creating a plug-in, for people who would like to work with me on their own projects (don’t hesitate to contact me using the social networks if you do!)

A circular approach to designing products

Product design is a topic that has been a real obsession of mine for some time now.

As a freelance developer, I’ve had a few opportunities to be involved in this stage of plug-in creation, though, for the most part, my clients already had a fairly clear vision of what they wanted and came to me to address a specific, identified need (such as emulating a particular electronic circuit, testing various algorithms for a specific type of effect, or troubleshooting an existing project).

What particularly interested me about this was seeing the extent to which completely unrelated fields could intertwine at this stage, ranging from graphic design, ergonomics, and UX design to researching best practices in existing products, and synthesizing the essential aspects of DSP blocks versus what should or could be made user-editable.

But it also involves elements like marketing, which can, or indeed must, be considered during the design phase rather than after the product is finished, specifically regarding market fit. This determines whether communicating about the product will be easy or difficult: what generates hype versus boredom; what is innovative and requires user education versus what is commonplace or immediately familiar; and what is currently trending versus what doesn’t yet exist in any form.

From there, we would look at the needs of musicians and future users, ranging from the needs they articulate to those they aren’t even aware of, whether these foster creativity or simply provide practical utility, and even consider what inspires or empowers the tool’s developers themselves (since those who spend their days creating plug-ins are often musicians or artists in their own right). We would also take into account the biases and blind spots shared by all these groups, their frustrations, and the underlying motivations for embarking on a new plug-in development project.

What fascinates me about this is that product design is a sort of circular process - one that integrates all these elements and fosters communication between them - yet it can also be a factor that negatively stands out in products that fail to achieve the expected success. Across all disciplines, we frequently see examples of products where questions regarding the viability of the product or service were raised only at a late stage, or indeed never asked at all.

Ever faster and for less time

It must be said that, in this year 2026, I personally observe that many things have changed in the world of plug-ins. To summarize, I would say that the advent of AI has caused a stir, but has also accelerated a number of things that already existed, such as market saturation or the continuous devaluation of plug-ins that are put online, driven in particular by the phenomenon of “promotional periods that are no longer the exception,” with these famous plug-ins officially on sale at 250 euros but whose average value over 365 days is less than 25% of this price, with a minimum of 5-6 different promotional periods per year (see for example these deals from the promotions page of the Synth Anatomy blog).

Furthermore, some of my clients (those who rely on human expertise) have faced mounting pressure to alter their development practices to ensure their products succeed in these peculiar times. This involves issues such as the time spent developing a plug-in versus the need to accelerate production to churn out ever more plug-ins, even if the resulting software offers fewer features and receives support for a shorter period. This battle for the attention of potential users is waged not only against direct industry competitors but also against other leisure activities beyond music creation, and even against the passive consumption of educational content that is rarely put into practice.

But above all, for me, it highlighted flaws in their product design approach, flaws that weren’t significant five years ago but cause concrete problems today. That is why I began educating myself on the subject and increasingly advising my clients on it over time, with a strong focus on monitoring industry best practices, particularly among a few key brands.

Create a new freeware plug-in in 2026?

When the KVR DC 2026 was announced, I have to admit I wasn’t keen on participating, given my work schedule at the time and the fact that I hadn’t yet fully visualized the release of my mystery commercial plug-in. I had a few ideas, none of which were feasible within three months, but certainly nothing that really grabbed me.

Then, while working on a Shimmer reverb algorithm (yes, the kind that feeds the reverb signal into a pitch shifter set an octave higher, popularized by Brian Eno, U2, and Valhalla’s plug-in), I found myself having fun trying to sustain the feedback loop for as long as possible. I added various effects that could be tweaked in real-time, like dynamic compression to keep the output from exploding, until it felt less like using a mere effects box and more like playing an actual instrument.

I was also significantly influenced by a few recent personal sonic adventures in electronic music, which I’ll have the chance to tell you more about later, that helped inspire, among others, a friend and ambient music composer, the artist Almost Random, who has just released a new album on Cyclical Dreams.

In any case, that was the moment it clicked for me; I felt the need to turn these experiments into something concrete, seizing the opportunity provided by the contest to also work on the elements I mentioned earlier. The project’s name came later, I liked its “French” ring, and I immediately envisioned moving between the Korg “Submarine” reverb (found in the Logue series) and the “Riser” and “Space” reverbs (specifically the -12 and +12 Shimmer settings, which sound fantastic), which also inspired the graphic design you’re familiar with. The trouble was, the first prototype I had on hand was nowhere near good enough for a commercial release, and it was already early June when I made these decisions. I had just one month to see the project through to completion, on top of my regular job!

Product Design & Automation

So, I kept working on the concept-focusing first on the sound, of course, but also on the controls I’d make available to the user. I spent time figuring out exactly what that #j$&@ Chaos control would do, it kept changing right up to the last minute, and addressing the low-frequency heavy content, which gave me trouble early on and needed to be tamed as much as possible. I then spent a few hours looking for graphic design inspiration in the hardware world, specifically Erica Synths pedals (which fit perfectly with their existing lineup of FX and instruments), but also drawing ideas from various delay and reverb units, and even other instruments, that I really like.

At the same time, I wanted something that wasn’t just another “Shimmer” reverb; I wanted to add a unique twist, which, as you’ve guessed, turned out to be the self-oscillation and drone generation features. I tried to make this as fun as possible, even though the workflow is unconventional, requiring you to regularly record the plug-in’s output rather than just relying on automation to get the best results, and I needed the GUI to reflect that aspect. I also figured this would help the plug-in stand out from recent market releases.

Some other aspects of the design I tried to focus on:

• I didn’t want my Shimmer effect to sound generic or even cliché
• I wanted something with personality and control, specifically over the Shimmer delay length, which allows for adjusting the build-up that is often fixed in this category of effects
• I took the time to determine which potentiometers should be highlighted, and considered factors such as their placement on the interface and the use of different sizes and concepts
• Various elements needed to encourage users to experiment with the drone and instrument sections of the plug-in, by eliminating friction with the UX, gently prompting them to explore, whether through built-in tutorials, color-coding, preset categories choice, and so on
• Finally, I wanted the sweet spots to be accessible with a minimum number of operations, so I had to merge controls or turn some of them into macros that affect multiple internal parameters.

And fortunately, I was very quickly very happy about the way it sounded, and it made me want to create interesting music with it, something I tried to demonstrate with the audio demos I released (with the help of Olivier VM).

Anyway, things started to take shape, and then I really had to nail down the layout, choosing the final controls, the interface arrangement, sizes and positions, and DSP interactions, as well as the graphics. I’m very confident about the coding side of things, having worked in this field for decades. This experience has given me both the know-how to create new things and a reusable codebase-covering elements like plug-in parameters, basic DSP blocks, and preset systems, that I’ve already automated through my previous freeware projects, allowing me to bypass AI entirely.

On the graphics side, however, the timing leading up to the deadline started to worry me a bit… I have solid skills in Affinity Photo, Knobman, and Blender 3D, but I’ve been procrastinating on redesigning the GUI for Spaceship Delay for years; I remember being frustrated when it was first released because I hadn’t had the time to fully realize my aesthetic vision. I wasn’t sure I could pull it off for Voyage Voyage.

I had actually started working on a project to automate plug-in design, which, unfortunately, wasn’t finished yet, as well as on categorizing an element I find fascinating: hardware knobs and, of course, their software counterparts when using a skeuomorphic aesthetic. I came across a pretty interesting article featuring the image you see below (can you guess which ones I used?).

Finally, and let’s be very clear about this right from the start, the graphic design for Voyage Voyage is the result of my work in Affinity Photo, involving layers of diverse materials and the use of the venerable Knobman3D for the potentiometers. I had a very personal vision for the design that evolved through iteration; I incorporated some reworked, royalty-free assets from the web, a drawing by my daughter that I vectorized, and some AI-generated abstract elements for the pedal’s background, something I recently discussed over at Bedroom Producers Blog. I also briefly tested the capabilities of certain paid AI models while I was learning the ropes. I quickly realized that the results I was achieving on my own were great, and way more satisfying in every respect, even with my current skill level!

While wrapping up the project, I took the opportunity to refine a few features I’d wanted to include in my plug-ins for a while: the resizing strategy, improvements to the preset system, and new Easter eggs, specifically one utilizing Roland Rabien’s JUCE module gin_controllers, which I reworked to support all types of joysticks, as well as a feature that simulates kicking a spring reverb unit!

I also took a look at JUCE’s accessibility features, integrated into the core modules since version 6.1 and compatible with Narrator on Windows and VoiceOver on macOS, by adding descriptions for each control and quickly testing how the plug-in operates with these tools. I also checked visual accessibility using the “Colour Simulations” tool on Windows. The resulting image for deuteranopia suggests I still have some work to do on my graphic design, but it will be completed on version 1.1.

So I’m very happy to have seen this plug-in project through to completion, regardless of the contest outcome, and I fully intend to keep talking about it and adding a few more features after the competition ends. I also hope you enjoyed my long post! If you liked my work and my explanations, please feel free to download the plug-in and vote for it, as well as for the work of the other developers participating in the KVR Audio Developer Challenge. You might also notice that I’ve done a bit of cleanup on my two websites (here and at musicalentropy.com).

Voyage Voyage by Musical Entropy

Stay safe!

• NTS-1 and NTS-1 mk2
• NTS-3
• Minilogue XD
• Prologue
• Drumlogue
• microKORG 2

Depending on the plug-in, this allows the units to gain capabilities in areas such as delay, reverb, modulation, and general effects-as well as synthesis and sound generation-thereby making the devices in question far more versatile very quickly.

For my first foray into this platform - which I’ve particularly loved ever since I had the chance to review the Korg NTS-1 for AudioFanzine, an experience that allowed me to dive into the world of associated third-party plug-ins - I chose to develop and release a somewhat unique type of effect: a phaser algorithm, called Phaser-Vibe.

I want to say that I particularly appreciate the effects and content Korg has offered in this regard. I’ve always had a soft spot for the reverbs found on each machine - such as “Submarine” or “Horror.” However, I wasn’t fully satisfied with the modulation options, especially the phasers - particularly given that I’ve had the opportunity to work on this subject over the years, notably while developing the Spaceship Delay plug-in.

There are actually various ways to design a digital phaser algorithm, and each type produces a very different sound. Fundamentally, a phaser is an effect that mixes a “dry” signal with a “wet” signal processed by several all-pass filters in series. These filters do not affect the signal’s amplitude-frequency response but do drastically alter its phase. This is how phasers operate in the analog world, and - most of the time - in the digital realm as well.

While this phase distortion is barely audible on its own, it becomes noticeable when the single parameter of these first-order all-pass filters-the cutoff frequency around which the phase distortion is centered - is modulated (for instance, by an LFO). Depending on the LFO’s waveform and the amplitude of the resulting pitch-shifting or vibrato effect (features sometimes found in commercial phasers), this creates a relatively subtle sound. However, the effect’s characteristic sound emerges primarily when this processed signal is summed with the dry signal: the phase distortions transform into clearly audible notches-especially when they are in motion.

It is also worth noting that these notches can be produced in various ways. Some digital phasers generate them directly-without relying on a mix between dry and wet signals-and allow for more advanced control over their spacing compared to the classic phaser design, which applies the same cutoff frequency to all all-pass filters, resulting in a somewhat predictable sound. Other phasers employ second-order all-pass filters, which further alters the characteristics of the effect. Then there are effects that create these notches using a short delay line, where the specific delay time determines the number of notches. This is the standard operating principle for chorus and flanger effects; in these cases, the notches are generally spaced linearly (with a constant frequency interval in Hz between them), whereas in phasers, they are spaced logarithmically (with a constant interval in octaves between them).

If we also add a feedback loop between the output and the input, these dips take on a different shape, and the effect becomes more pronounced-even taking on a somewhat metallic timbre.

So, what actually makes my phaser different from others? For starters, it features a very classic structure with a feedback loop, but the cutoff frequencies differ across the all-pass filters. This is actually quite unusual for analog phasers-except for Uni-Vibe clones or units influenced by them. Historically, those were the first types of phasers-originally designed to emulate Leslie cabinets in pedal form-and that specific characteristic (which gave my effect its name, unrelated to “vibe-coding”) contributes to the sound of that legendary pedal (along with other factors, of course).

Another point of interest is the shape of the modulation signal from the LFO and how it translates into an actual cutoff frequency for the all-pass filters. This might seem insignificant at first, but factors such as:

• the transition from a linear to an exponential curve,
• the inertia of certain components - like the optocouplers acting on the filters - which itself depends on the modulation depth or polarity,
• the initial waveform shape,

all contribute to the phaser’s final audible signature; indeed, some analog phasers (and their emulations) derive their distinctive coloration from these characteristics, rather than simply from the number of notches.

Finally, let’s add the modeling of all the signal filtering elements within the phaser beyond just the all-pass filters, such as the feedback path and the tendency of certain units (like the MXR Phase 90) to saturate in a specific way. Combine this with the algorithm’s ability to accurately model feedback-specifically using the renowned “zero-delay feedback” techniques I’ve employed here instead of a simple delay-and you get a range of distinct tonal characters, including the one I’m presenting here!

Incidentally, if you want to use it on something other than a Korg machine, it happens to be available in Spaceship Delay as well, and can be used standalone.

Fun fact: there is also a version of my phaser algorithm in the Phazerville Suite firmware for my favorite Eurorack module, the Ornament & Crimes - specifically the Teensy 4.1 version like the After Later Audio Ornate Criminal, which features more I/O, new audio effects, more power, and more of everything. If you’re into modular synths, be sure to check it out too!

The Great Escape by Musical Entropy

The plug-in was therefore created at the end of March 2020, during the infamous COVID lockdown. I had seen a few initiatives from artists and developers who wanted to communicate the importance of respecting the lockdown and who were offering ways to keep people busy and make it more fun, and I thought the idea sounded very interesting. So I decided to do something along those lines and developed everything in just a few days, without any vibe coding :)

Moreover, a few weeks earlier, I had the chance to participate in a show in Paris, called Festival United Guitars 2020, organized by Ludovic Egraz, editor-in-chief of Guitare Xtreme magazine, and leader of the musical project of the same name United Guitars which showcases compositions produced by some of the best French guitarists.

At that show, I had the chance to try out a few pedals, and I was particularly fond of the famous The Great Escape by Thrilltone. This tremolo sounded really good, and using dynamics to manipulate its parameters with just a few knobs struck me as a brilliant idea, and a fun way to rediscover playing with a tremolo. Just a few days later, I loved the concept so much that I tried to recreate it in a plug-in, mostly by ear, especially the dynamics side. And that resulted in this magnificent prototype :)

When I decided to release a new freeware, the idea of completing this modeling work seemed very good to me, especially since I had the opportunity to discuss with Pierre-Benoît Prud’homme, the founder of the pedal and the Thrilltone brand, which also makes great distortion pedals + others, whose graphic aesthetic I greatly appreciate (just look at this beauty):

So we got in touch, and the idea of releasing a paid software tremolo emulation, initially with a rough model in a free version, seemed like a good one. The freeware plug-in was released as a VST/VST3/AU/AAX and even a Linux version, a first for me at the time. The plug-in regularly displayed lockdown recommendations, which I removed in a later version, and used the original pedal graphical designs, generously provided by Pierre-Benoît.

The idea was to discuss collaboration again for the paid version later, and to properly promote the original pedal as well, which apparently sold a bit better thanks to me! We’ll also skip over the fact that the freeware included an Easter Egg that’s still there and is quite entertaining in itself :)

I’ll mention again that this is an analog tremolo emulation, with its subtle touch of signal coloration even when the tremolo isn’t active. Beyond its sounds, it distinguishes itself from a classic tremolo by its ability to easily add variety to the effect itself, using a parametric envelope follower that can vary the speed and amount of the tremolo based on the signal’s dynamics.

The whole thing is also very well explained in the original manual, in the educational aspect of my plug-in which I had particularly pushed, and applies equally well to guitars as to vocals or synthesizers sounds.

Then, very recently, and as you know, I released a new version of the plug-in.

This project allowed me to incorporate the latest updates to the JUCE framework and my codebase, properly implement support for the latest operating systems, add CLAP to the list of available formats (currently only on Linux), and to refine the emulation to better match the behavior of the original pedal, which I now own. However, paid collaboration is unlikely, and for this reason I’ve focused on creating an emulation that closely resembles, but isn’t strictly identical to the original, with a few improvements over the initial version I released.

In any case, I’m very proud of this plug-in, very happy to have been able to work with Pierre-Benoît on the emulation of his awesome software tremolo, and I’ll keep you updated on my upcoming releases and work. I’ll have some announcements to make very soon about that :)

Stay safe and stay tuned!

I decided to get back to work on my Musical Entropy plug-ins, after spending a few years focusing exclusively on my freelance developer career. As you may have noticed, I have just released a new version of my freeware plug-in Spaceship Delay, which was essentially an opportunity to relaunch the project and clean up all the source code. I did the same with The Great Escape, and this allowed me to make available the latest versions of these projects, with bonus installers, and all the classic plug-in formats under Linux. You can find them in the usual locations here or on my website, always free and accessible without any kind of registration process. By the way, I will also mention The Great Escape in a separate post, as I have not yet had the honor of talking about it here.

So Spaceship Delay has had its fair share of updates, mainly to ensure compatibility with the latest OSes, fix a few bugs here and there, improve the usability and visibility of the embedded tutorial section, add resizing, a dark theme, improved the phaser algorithm with a zero-delay feedback structure (used in my recent design for the Eurorack module Ornaments & Crime alternate firmware Phazerville as well), and optimize the execution of certain parts of the code, such as the convolution spring reverb (it’s an improved version of the code I sold a few years ago to JUCE for the DSP module). Over a hundred fixes later, I find the plug-in much more stable, it does exactly what it’s supposed to do without any drift (which wasn’t the case before, lol), and it’s therefore ready for me to add cool new features.

What kind of new features you’re asking? A new, modern design for the user interface obviously, but also things that will make Spaceship Delay even more unique and true to the vision I had when I first released it at the same time, as well as a synthesis of everything I’ve worked on since then, and all the things I love in delay FXs now that I’m a bit more involved in music production in general than before.

Stay tuned!

I even spent one week in FXpansion headquarters to work in a hellish but fun pace ! And I had a great time hanging out with the JUCE + FXpansion guys. Basically, I provided some code to help JUCE users writing plug-ins who need a convolution engine, filter design and processing functions, oversampling, some fast approximation of functions…

I have written this blog message today because recently I have create a few topics on the JUCE forum related with my work, and I wanted to post links to them there :

• DSP module discussion / Structure of audio plug-ins API
• DSP module discussion / New Oversampling class
• DSP module discussion / New AudioBlock class
• DSP module discussion / IIR::Filter and StateVariableFilter
• DSP module discussion / Fast function computation classes
• DSP module discussion / New class SIMDRegister
• DSP module discussion / New classes in the maths folder

Otherwise, I’ll be again at ADC 17 this year in London, presenting a talk called Fifty Shades of Distortion (which will be on Youtube a few weeks later), and I’ll update Spaceship Delay when I’m back with a few improvements + new functionalities!

They started a partnership with the Goldsmith University in London and the IRCAM in Paris, about their european project called RapidMix for Realtime Adaptive Prototyping for Industrial Design of Multimodal Interactive eXpressive Technology (yes I know it’s long). To summarize, it’s a Machine Learning C++ toolkit/API, made for giving a fast and easy access to Machine Learning algorithms to developers and artists, in order to create innovative human-computer interfaces, for gesture recognition and creation of new musical instruments.

So, two weeks before the hackathon, the JUCE team asked me to develop a JUCE module wrapping a section of RapidMix called RapidLib, created by Michael Zbyszyński. It is the little brother of the Wekinator, designed by Rebecca Fiebrink also at Goldsmith. You might know them for the amazing talks they did at the Audio Developer Conference 2016 in November, and for the MOOC Machine Learning for Musicians and Artists. I did code the module, some JUCE examples, and I did also a talk at the beginning of the Hackathon so everybody could grab the basic concepts and start doing something cool.

You can see the slides I have used during my talk here : Hackathon Slides.

Machine Learning ?

What’s interesting here is that I didn’t know anything at all or nearly about Machine Learning before! But I learnt everything I could about the subject in that short amount of time in order to do the job successfully, and to be able to explain how to use the JUCE module and do cool things with it. Ultimately, everything worked as expected, and I also got nice reviews about my work and my talk from the people involved in the RapidMix project, for which I am very proud!

Anyway, the reason I’m talking about that is to tell you how much I am excited about this collaboration and the things that are going to happen next. The JUCE module isn’t available to public yet, since it is in a very early alpha version right now, but it will be released in the future, and the people who have been able to have fun with it already in the Hackathon were very happy to catch this opportunity too.

Then, why am I so excited about this ? Since I’m new in this area, I wouldn’t be able to explain in details what is Machine Learning and how it works, but here is what I have learnt already (you can have a look for my talk slides too) :

• The very principle of Machine Learning is the use of an algorithm, which is trained with some data provided by the user. The training set of data is made of samples with inputs + associated outputs. This algorithm then is expected to have a given behaviour, when receiving some random input data.

• There are two main kinds of Machine Learning algorithms, ones to do classification, and ones to do regression.

• In classification applications, the output of the algorithm is an integer variable or a class label. For example, a classification algorithm is trained with pictures of animals, and the associated animal names. When a new picture is given to it, the algorithm is supposed to be able to tell you what animal is in the picture.

• In regression applications however, the outputs are continuous, they can be any float/double variables. As an example of regression, imagine you want to model a mathematical function with a Machine Learning algorithm, like an hyperbolic tangent. You train your regression algorithm with a set of input + output data that you calculate yourself. Then the algorithm processes any input value and returns an output which is supposed to be as close as possible of the original mathematical function result.

• In order to train properly Machine Learning algorithms, it is common to use features, which means sampling a large amount of data and compressing it or extracting some meaningful information from it. For example in audio, it is possible to train an algorithm with a RMS value taken from a 200 ms buffer of audio samples, instead of feeding the algorithm directly with that buffer. It is also possible to extract some information such as the pitch, the frequency response etc. With a video signal, you can reduce the resolution of the image or extract a general brightness value.

Applications of Machine Learning

Then, what can we do from all of this stuff ? In the audio examples I saw, regression algorithms were used to map and interpolate parameters of an audio synthesizer using controllers such as the computer mouse, a Leap Motion, some ROLI Blocks, a joystick etc. In her videos, Rebecca Fiebrink uses a lot of weird controllers and the Wekinator to create new instruments with new ways of interacting with a computer and making music. It is fun also to see that the musician can be involved during the training phase, and it is not all the time the duty of the application developer to train the Machine Learning algorithm.

During the Hackathon, I coded something cool too (but I wasn’t able to complete it before the end of the event unfortunately). I developed a drum machine application which can play 3 samples (a kick, a snare and a hi-hat sample), with 3 patterns that can be modified by the user with its mouse, or live if he connects a ROLI Blocks to its computer in USB. Then, he can “train” a classification algorithm with the microphone input of its computer, in order to associate outside sounds to one of the three patterns, using audio features extraction. This way, he can play the pattern 1 using a low-pitched sound for example, the pattern 2 with a white noise like impulse, the pattern 3 with anything else, whatever the user chooses. And finally, there is a simple delay audio effect with 4 parameters (delay time, feedback, mix, lowpass filter frequency) which can be “trained” with the mouse cursor. That means the user can associate the mouse cursor position with a given set of the 4 parameters, and then play with the audio effect simply by dragging the mouse over the window, which will move continuously its parameters like in the Michael Zbyszyński’s talk. Sounds fun isn’t it? I promise, I’ll make it available in a way or another at some point when completed.

The other participants did cool things too, even if it was difficult in a short amount of time to discover a new tool, a new technology, and to do something with them… But the beauty of JUCE is that it is possible to create a new project, handle all the dependencies of everything, and start coding in a few seconds! The JUCE module during the hackathon featured simple classification and regression classes, but unfortunately nothing yet on the feature extraction side, which was the main concern of the participants at the time. We all had to code some on our own, which explains partly why I wasn’t able to finish my application…

Conclusion

Anyway, there is a JUCE module RapidMix being developed right now. Machine Learning is a hot topic nowadays thanks to all the communication about Deep Learning, or about some current applications related with audio such as speech recognition or automatic composition of music. I’m still discovering the topic right now, but in my opinion having something as simple as a JUCE module to let audio developers have fun with Machine Learning, thanks to the work of the RapidMix team, could be something huge for musicians and sound engineers. Not because Machine Learning would help creating really innovative stuff, I mean all the applications that I have talked about could be done without it. But because a lot of complicated problems can become very simple to solve thanks to Machine Learning, and because these algorithms allow to experiment a lot of things very quickly! Finally, I think that current users of Machine Learning APIs / algorithms right now are most of the time in the academic world, and having something like a module in the JUCE library could make these features available easily to new people, with new application ideas, which is awesome.

I’d like to say also that I’m curious about the next developments in RapidMix, to see how it will perform compared with other Machine Learning APIs for C++. And I would love to add in Spaceship Delay some Machine Learning features I have in mind…

Anyway, thanks to all the people involved in this hackathon for organizing it, it was great to talk with them and with the participants back then!

Bibliography

If you want to learn more about Machine Learning, there are tons of information sources on the internet, but I suggest you to have a look there in priority :

• Machine Learning Crash Course Part 1, by Daniel Geng and Shannon Shih
• Machine Learning for Musicians and Artists
• The classic Andrew Ng’s Machine Learning course on Coursera

So the winners are Youlean with its Youlean Loudness Meter and Ursa DSP / Dave Elton with its delay Lagrange. It’s funny to see two delay plug-ins in the top 3!! Youlean Loudness Meter is probably one of the best loudness meter that you can find for free, and I have seen a video comparing it with commercial alternatives. Lagrange is probably a less versatile delay than mine, but it is using granular methods to get very interesting and singular sounds that cannot be obtained with other methods, close to what can be got with stutter effects or some specific flanger / chorus effects. I suggest you to try them if you have not yet.

I would like also to tell you that I’m very happy about having a blog post about Spaceship Delay in Peter Kirn’s blog, Create Digital Music, that I read a lot for a long time. The article is here : Spaceship Delay is an insane free plug-in inspired by hardware

And last thing, I have just updated Spaceship Delay to version 1.0.5. The new version includes a few things people wanted for a long time : mono/stereo versions of the plug-in for Pro Tools, Logic Pro X etc., a selector for the location of the Post-FX (before or after the mix control), a new tremolo Post FX and extra modulation options applied to the low-pass filters. There are also new presets and some extra tutorial content, so I suggest you to update your presets / tutorial folders as well.

That’s it for today. I’ll probably talk again about Spaceship Delay very soon since I have big plans for it in the future!

Low/High Cut and Low/High Shelf

These filters are more or less the standard 2nd order filters or “biquads”, modeling the analog circuit called State Variable Filter (SVF) behaviour, with a -12 dB/octave attenuation. The equations are derived from the famous Robert-Bristow Johnson EQ Audio Cookbook that every DSP engineer in the world knows, and uses sometimes without even knowing. Same for most of filter algorithms available in commercial software.

However, the original implementation does have some drawbacks well known, that I have been talking about in my recent Audio Developer Conference talk. The two main drawbacks applies to the behaviour of the filters in the high frequency range, and when the cutoff frequency is modulated fast. In the first case, a lowpass filter for example is attenuating way too much the high frequencies close to half of the sampling rate. In the second case, a quick modulation with the original implementation, using a simulation structure called Direct Form produces high amplitude artefacts at modulation.

In Spaceship Delay, I have solved these issues by using another simulation structure called Topology-Preserving Transform, which is well documented on KVR forums (DSP section) and in the free e-book of Native Instruments Vadim Zavalishin, The Art of V.A. Filter Design. I have also oversampled the filtering algorithm, since the section can produce some nonlinear stuff as we will see with the other kinds of filters. That gives me the right behaviour in high frequencies and when a parameter is changed by the user with automation. It will be even more useful if I decide to add in the modulation section in the future an envelope follower or a LFO modulating the cutoff frequencies. Moreover, this structure is already used in the phaser algorithm, for obvious reasons.

I’d like to say there is absolutely nothing really innovative there, since this kind of filtering algorithm can already be found in most of current commercial plug-ins involving filtering. The Low/High Cut feature in Spaceship Delay has nothing groundbreaking either, and that’s something that we find in most of other delay plug-ins, so its presence here is more than relevant without being something particularly new. However, I have not seen a lot of times some low and high shelf filters in this context, and I thought it would be a nice addition to drive further the nonlinear sections, put the delay line into resonant feedback in an interesting way, or simply to cut some frequency content like with the Low/High Cut filters in a different way. It’s my friend François-Maxime from Les Liens du Son who suggested me this to do this and he was right!

Japanese

The Japanese filter is obviously my take on the MS-20 / Monotron filters. You can find a very interesting study of its behaviour on Tim Stinchcombe’s website, that I have been reading a lot during the development.

Tim Stinchcombe study of Korg MS-20 filter

I’m not that proud of the realism and simulation quality of the result, since I’m still new in synthesizer filter modeling. But I think that the “japanse filter” in Spaceship Delay sounds quite good and is very interesting when put in a delay line, to get that screaming oscillating feedbacks everybody loves in “analog” delay software/hardware. As I said before, doing a delay and using that filter algorithm was the starting point of Spaceship Delay in terms of DSP. I wanted something that sounds a little like my Korg Monotron Delay, that I could use as a plug-in. I thought also that using other kinds of “Virtual Analog” filters in the delay line should be interesting too. So I did some research, trying also to dismiss the too obvious choices for adding a extra V.A. filter in the plug-in, and then I ended up doing modeling of the filter I called “German/Canadian”.

German/Canadian

For those who have read the embedded tutorial in Spaceship Delay, or seen the other blog posts, you already know the true identity of that filter. It is the one that can be found in the Meeblip Anode and Triode synthesizers, made by Peter Kirn from Create Digital Music and James Grahame from Blipsonic / Meeblip.

I have spent a lot of time studying the schematic of the filter that is available on GitHub, covered by a permissive Creative Commons and GPLv3 license, since the Meeblip hardware + software is open source! I wanted a filter in Spaceship Delay which would have second order attenuation like the Korg MS-20 filter (and not like the famous ladder filters), and the result sounded surprinsingly good too when put in a delay line.

My implementation, like for the Korg MS-20 filter, isn’t that realistic yet, since I used a simplified model, and because I have not reproduced yet the same mapping for the cutoff and resonance knobs than the one in the original units. However, my model is based on the equations of the original circuit, that I have studied first in LTSpice and then in a simulation context to make it sound as good as possible.

As you can see, for people knowing how to read a synthesizer filter schematic, it’s a filter looking more or less like a Twin-T VCF, acting like a lowpass filter, but removing also a little in the bass frequency range, giving it a very interesting sound signature in my opinion. It is possible to study it further by determinating its transfer function from the electronic equations :

I’ve made it zero delay feedback, but the way it saturates is still far from the original in my opinion, so I’m going to improve my model over the next weeks, and I’ll probably start by updating the mapping of the controls so it behaves like the original at least in a strictly linear sense.

As a side note, don’t forget to get the last version of Spaceship Delay to experiment with this filter, since the algorithm has changed a little since the beginning of the KVR Developer Challenge, to solve a few issues I had.

Next steps for Spaceship Delay

I got already tons of very nice comments, and feature requests for Spaceship Delay. I even got a very nice new skin proposal from a user of the KVR forums. I’ll probably spend a lot of time after the voting period to add the most interesting features people have submitted, and a few things I had also in mind and that I have not been able to finish before the deadline. But, what I can already tell you, is that I’m already working on the AAX version of the plug-in, and it should be available in a few days!

In the new version, I have solved a few bugs, and I have changed a little the way you can install it, to answer some user calls. Now, you can just put the data folder in the same place than the plug-in itself, so it will be easier to start playing with it for newcomers.

I’d like also to recall that Spaceship Delay has now a few presets, and a tutoriel system I’m very proud of, with tips and tricks embedded in the application, available from the “about tab” by clicking on Spaceship Delay logo. It looks like this:

I hope you’ll enjoy Spaceship Delay, and again, if you are a KVR forum member, don’t hesitate to vote for me if you like the plug-in!

When I was working on the modeling of the Echocord Super 76, I spent a lot of time doing measurement, and I have only tried to do something from them a few days before the KVR Developer Challenge deadline, to include the spring reverb impulse response in my plug-in. However, after the first calculations, I got something really strange as a result…

I was scared because I thought that I did something wrong with the measurement itself, like having a feedback loop in the recording that I didn’t saw at first, and I thought all my recordings were screwed up. You can imagine what was the impact of this a few days before deadline, with no way to redo easily the recordings since I don’t own the device. Fortunately, it turns out that there was a bug in my deconvolution algorithm, and so everything was finally fine!

However, I tried to re use the “wrong” impulse response again a few days ago, and I thought that the sound I got is somehow really cool, so I thought it might be a good idea to share it with you! I have added some artificial decay on the reverb tail so it doesn’t stop too suddenly. To use it, you need to open your favorite convolution reverb plug-in, and to load the impulse response inside. You can find it here :

Impulse Response

It happens a lot of times during development to have some kind of “interesting accidents”, when a bug or an unexpected thing happens in the code, and gives interesting sonic results…

Enjoy!