The author finds, as many do, that naive or first-approximation approaches fail within certain constraints and that more complex methods are necessary to achieve simplicity. He finds, as I have, that perceptual and spectral domains are a better space to work in for things that are perceptual and spectral than in the raw data.
What I don't see him get to (might be the next blog post, IDK), is getting into constraints in the use of color - everything is in 'rainbow town' as we say, and it's there that things get chewy.
I'm personally not a fan of emissive green LED light in social spaces. I think it looks terrible and makes people look terrible. Just a personal thing, but putting it into practice with these sorts of systems is challenging as it results in spectral discontinuities and immediately requires the use of more sophisticated color systems.
I'm also about maximum restraint in these systems - if they have flashy tricks, I feel they should do them very very rarely and instead have durational and/or stochastic behavior that keeps a lot in reserve and rewards closer inspection.
I put all this stuff into practice in a permanent audio-reactive LED installation at a food hall/ nightclub in Boulder: https://hardwork.party/rosetta-hall-2019/
I really like your LED installation in Rosetta Hall, it looks beautiful!
https://en.wikipedia.org/wiki/Wicked_problem
Kinda funny but I am a fan of green LED light to supplement natural light on hot summer days. I can feel the radiant heat from LED lights on my bare skin and since the human eye is most sensitive to green light I feel the most comfortable with my LED strip set to (0,255,0)
(Note both the scanner in front of KITT and the visual FX on his dashboard when he speaks, which changes from season to season.)
When the author says:
> Every commercial audio reactive LED strip I've seen does this badly. They use simple volume detection or naive FFTs and call it a day. They don't model human perception on either side, which is why they all look the same.
well no, if they sell, then they are doing just fine until someone comes up with the $next $thing
The wickedness comes from wanting something that works just as well for John Summit as the Grateful Dead as Mozart and Bad Bunny.
But it seems like you could cheat for installations where the type of music is known and go from there. The other cheat is to have a "tap" button, and to pull that data and go from there.
mental note: the thought "it can't be that hard" when obviously it is sent me down a rabbit hole for a couple of hours
I don't think FFTs are particularly good for music visualization if you're trying to be expressive because there isn't a particularly meaningful mapping from an FFT to the subjective experience of music and it adds significant latency. I ended up using a few stacked bandpass filters as well as mixing the raw PCM into the light strip for texture. Compressor with a slow attack and even slower release for auto-leveling (the release has to be like 20 seconds to make sure you don't up the gain a bunch during a breakdown in the music). I ran all the realtime stuff on one core of an ESP32 and a bluetooth stack and all the UI stuff on the other. I was getting about 200FPS on a strip of about 120 SK9822s w/ a custom HDR driver giving me about 11.5 bits of color per channel.
I really miss my bike, watch your shit on caltrain.
There's so many massively better solutions than FFTs when it comes to the way people perceive sound.
And of course, by the time I got it to work perfectly I never looked at it again. As is tradition.
I wonder if transformer tech is close to achieving real-time audio decoding, where you can split a track into it's component instruments, and light show off of that. Think those fancy Christmas time front yard light shows as opposed to random colors kind of blinking with what maybe is a beat.
There was a nice paper with an overview last year too https://arxiv.org/html/2511.13146v1 that introduced RT-STT which is still being tweaked and built upon in the MSS scene
The high quality ones like MDXNet and Demucs usually have at least several seconds of latency though, but for something like displaying visuals high quality is not really needed and the real time approaches should be fine.
At the end it's "just" chunking streamed audio into windows and predicting which LEDs a window should activate. One can build a complex non-realtime pipeline, generate high-quality training data with it, and then train a much smaller model (maybe even an MLP) with it to predict just this task.
Another related project that builds on a similar foundation: https://github.com/ledfx/ledfx
It was fiddly, and probably too inaccurate for a modern audience but I can't claim it was diabolically hard. Tuning was a faff but we were more willing to sit and tweak resistor and capacitor values then.
“Most people who attempt audio reactive LED strips end up somewhere around here, with a naive FFT method. It works well enough on a screen, where you have millions of pixels and can display a full spectrogram with plenty of room for detail. But on 144 LEDs, the limitations are brutal. On an LED strip, you can't afford to "waste" any pixels and the features you display need to be more perceptually meaningful.”
(And it looks like the 7 frequencies are not distributed linearly—perhaps closer to the mel scale.)
I tried using one of the FFT libraries on the Arduino directly but had no luck. The MSGEQ7 chip is nice.
I remember thinking really hard on what to do with color. Except like you say mine is pretty much a naive fft.
https://github.com/aleksiy325/PiSpectrumHoodie?tab=readme-ov...
Thanks for reminding me.
- The more filters I added the worse it got. A simple EMA with smoothing gave the best results. Although, your pipeline looks way better than what I came up with!
- I ended up using the Teensy 4.0 which let me do real time FFT and post processing in less than 10ms (I want to say it was ~1ms but I can't recall; it's been a while). If anyone goes down this path I'd heavily recommend checking out the teensy. It removes the need for a raspi or computer. Plus, Paul is an absolute genius and his work is beyond amazing [1].
- I started out with non-addressable LEDs also. I attempted to switch to WS2812's as well, but couldn't find a decent algorithm to make it look good. Yours came out really well! Kudos.
- Putting the leds inside of an LED strip diffuser channel made the biggest difference. I spent so long trying to smooth it out getting it to look good when a simple diffuser was all I needed (I love the paper diffuser you made).
RE: What's Still Missing: I came to a similar conclusion as well. Manually programmed animation sequences are unparalleled. I worked as a stagehand in college and saw what went into their shows. It was insane. I think the only way to have that same WOW factor is via pre-processing. I worked on this before AI was feasible, but if I were to take another stab at it I would attempt to do it with something like TinyML. I don't think real time is possible with this approach. Although, maybe you could buffer the audio with a slight delay? I know what I'll be doing this weekend... lol.
Again, great work. To those who also go down this rabbit hole: good luck.
But perhaps you'd get better results if more of a ML speech/audio recognition pipeline were included?
Eg. the pipeline could separate out drum beats from piano notes, and present them differently in the visualization?
An autoencoder network trained to minimize perceptual reconstruction loss would probably have the most 'interesting' information at the bottleneck, so that's the layer I'd feed into my LED strip.
Effects themselves are written in embedded Javascript and can be layered a bit like photoshop. Currently it only supports driving nanoleaf and wled fixtures, though wled gives you a huge range of options. The effect language is fully exposed so you can easily write your own effects against the real-time audio signals.
It isn't open source though, and still needs better onboarding and tutorials. Currently it's completely free, haven't really decided on if I want to bother trying to monetize any of it. If I were to it would probably just be for DMX and maybe midi support. Or maybe just for an ecosystem of portable hardware.
This allowed the device to count the beats, and since most modern EDM music is 4/4 that means you can trigger effects every time something "changes" in the music after synching once.
The classic "Color Organ" from the 70's.
I tried recreating the app (and I can connect via BT to the lights) but writing the audio-reactive code was the hardest part (and I still haven't managed to figure out a good rule of thumb or something). I mainly use it when listening to EDM or club music, so it's always a classic 4/4 110-130bpm signature, yet it's hard to have the lights react on beat.
To solve this I tried pre-processing the audio, which only works with recordings obviously. I extract the beats and the chords (using Chordify). I made a basic animation and pulsed the lights to the beat, and mapped the chords to different color palettes.
Some friends and I rushed it to put it together as a Burning Man art project and it wasn't perfect, but by the time we launched it felt a lot closer to what I'd imagined. Here's a grainy video of it working at Burning Man: https://www.youtube.com/watch?v=sXVZhv_Xi0I
It works pretty well with most songs that you pick. Just saying there's another way to go somewhere between (1) fully reactive to live audio, and (2) hand designed animations.
I don't think there's an easy bridge to make it work with live audio though unfortunately.
I haven't seen that done yet. I think it's one of those Dryland myths.
I think its more likely going to come from a direct integration with existing synthesis methods, but .. I’m kind of biased when it comes to audio and light synthesizers, having made a few of each…
We have addressed this expert tuning issue with the MagicShifter, which is a product not quite competing with the OP’s work, but very much aligned with it[1]:
.. which is a very fun little light synthesizer capable of POV rendering, in-air text effects, light sequencer programming, MIDI, and so on .. plus, has a 6dof sensor enabling some degree of magnetometers, accelerometers, touch-sensing and so on .. so you can use it for a lot of great things. We have a mode “BEAT” that you can place on a speaker and get reactive LED strips of a form (quite functional) pretty much micro-mechanically, as in: through the case and thus the sensor, not an ADAC, not processing audio - but the levers in between the sensor and the audio source. So - not quite the same, but functionally equivalent in the long-rung (plus the magicshifter is battery powered and pocketable, and you can paint your own POV images and so on, but .. whatever..)
The thing is, the limits: yes, there are limits - but like all instruments you need to tune to/from/with those limits. It’s not so much that achieving perfect audio reactive LED’s is diabolically hard, but rather making aesthetically/functionally relevant decisions about when to accept those limits requires a bit of gumption.
Humans can be very forgiving with LED/light-based interfaces, if you stack things right. The aesthetics of the thing can go a long way towards providing a great user experience .. and in fact, is important to giving it.
[1] - (okay, you can power a few meters of LED strips with a single MagicShifter, so maybe it is ‘competition’, but whatever..)
I get a cert mismatch on that site, and when clicking the shop link I end up at https://hackerspaceshop.com/ which is advertising an online fax service.
That can be done with analog electronics, but even half an analog vocoder needs a lot of parts. It's going to be cheaper and more reliable to simulate it in software. This uses entirely IIR filters, which are computationally cheap and calculated one sample at a time, so they have the minimum possible latency. I'd be curious if any LLM actually recognizes that an audio visualizer is half a vocoder instead of jumping straight to the obvious (and higher latency) FFT approach.
There’s plenty of visual experiments of pianists doing this “rock band” “guitar hero” style visualization of notes.
The hardest part IMO is distributing it to a lot of LEDs over a large area. It usually involves finding the single serial maximum power length, cutting a bit before that, and SPI multiplexer(s) like SP901E, and strategically distributing right-sized power supplies. SPI amplifiers are also sometimes needed on long runs that skip across areas.
Edit: Oh wait, that project needs a PC or Raspberry PI for audio processing. WLED does everything on the ESP32.
And yea, I agree with the article. In my past I've also dabbled in audioreactive for LEDs and it's fiendishly difficult to make anything interesting.
Make it react too much, and it's chaos, and inversely when the algorithm reacts less the audio, it's boring.
And in all cases it's really not easy to see what the leds are doing in correspondence to all the complexity of music.
For my use case I want something fully portable and battery powered anyways. So the audio stuff should happen on the ESP32. (Or on my phone, that might work too)
Everything is relative, though. In terms of maximums, a Pi 4 (for example) can use up to about 7 Watts under load by itself, which adds up fast when operating on batteries.
But a single 1 meter string of 144 WS2812B LEDs can suck down up to around 43 Watts, and 43 is a lot more than 7. :)
Lighting rigs are thirsty. The processing (even if it's the whole Pi) is generally a small drop in the bucket.
In short, audio and visual perception do not map perfectly. Humans don't have a linear perception of either so a perfect A to D then D to A conversion yields unsatisfying results.
This kind of visualisation is an arbitrary artistic choice, not just a function of inputs. A can of spray paint is a tool that needs to be mastered, and it's different from, say, oil paints and brush. LED strip is just another tool. You need to figure out first which movements, pulses, patterns it can produce, and what “looks good”. Those would be the strokes.
The same happens on the other side. Choosing how to interpret sound is also an artistic choice. Everyone does the audio spectrum because everyone has seen the audio spectrum, and considers it a “natural” projection to some one-dimensional form. It only seems “natural” because of all of the graphs you've seen in textbooks. The need to use log scale or smoothing when real audio is not a pure set of harmonics is how “nature” has to smuggle itself back into the abstract reasoning. Beats work for a reason: what we call “modern music” is defined by its constant use of rhythm. When you have a different kind of sound, you need to process it differently.
So the goal is to match something you hear in the audio with something nice that the LED strip does. Which is also an arbitrary artistic choice, and can only be judged as a whole. There is no rule that tone has to match specific position or specific colour. Also, people rarely look at LED strips on their own. Just like film crews, you need to take ambient environment into account, and sometimes increase contrast with light, sometimes blend everything together. Some kind of compressor/expander for dynamic range is probably needed for different environments.
Often the thing that reflects the light is more important. I'd even say that the best way to increase the complexity of that low resolution source is to combine it with some complex object instead of using just the straight line. A Christmas tree should come to mind as an example.
It is wrong to think that the goal of such projects is to figure out a perfect simple process that turns one array of values into another. Their goal is to make people feel something.