I like this writeup but I feel like the title doesn't really tell you what it's about ... to me it's about creativity within constraints.
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 didn't go into much detail about it but there's a whole rabbit hole of color theory and color models. For example, the spectrum effect assigns different colors to different frequency bins, but also adjusts the assignment over time to avoid a static looking effect. It does this by rotating a "color angle" kind of like the HSL model.
I really like your LED installation in Rosetta Hall, it looks beautiful!
doctorhandshake 34 minutes ago [-]
Thanks! Great article - would like to read one about the color rabbit hole pls ;)
PaulHoule 1 hours ago [-]
Yeah, "diabolical" overstates it. It isn't a 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)
scottlawson 1 hours ago [-]
I'd actually argue it has some wicked problem characteristics. The input space is enormous (all possible audio), perception is subjective and nonlinear, and there's no objective function to optimize against, only "does this feel right?". Every solution you try reframes what "good" means. It's not as hard as social planning but is way harder than it sounds, no pun intended.
(Note both the scanner in front of KITT and the visual FX on his dashboard when he speaks, which changes from season to season.)
fragmede 6 minutes ago [-]
fta: The biggest unsolved problem is making it work well on all kinds of music.
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.
copypaper 8 minutes ago [-]
This is awesome! I did a similar project in college for one of my classes and ran into the same exact walls as you.
- 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.
The real killer is that humans don't hear frequencies, they hear instruments, which are a stack of frequencies that roughly sometimes correlate with a frequency range.
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.
adzm 7 minutes ago [-]
real time audio stem separation is already possible, some specific models can even get around 20ms latency (HS-TasNet) https://github.com/lucidrains/HS-TasNet
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.
menno-dot-ai 2 hours ago [-]
Woow, this was my first hardware project right around the time it released! I remember stapling a bunch of LED strips around our common room and creating a case for the pi + power supply by drilling a bunch of ventilation + cable holes in a wooden box.
And of course, by the time I got it to work perfectly I never looked at it again. As is tradition.
scottlawson 1 hours ago [-]
That's awesome to hear! Sometimes the journey is the destination, its a great project to get started with electronics.
Always been very interested in audio-reactive led strips or led bulbs, I've been using a Windows app to control my LIFX lights for years but lately it hasn't been maintained and it won't connect to my lights anymore.
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.
JKCalhoun 3 hours ago [-]
I made a decent audio visualizer using the MSGEQ7 [1]. It buckets a count for seven audio frequency ranges—an Arduino would poll on every loop. It looks like the MSGEQ7 is not a standard part any longer unfortunately.
(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.
Have you ever seen anything like a MSGEQ14 or equivalent? It would be cool to go beyond 7 in such a simple-to-use chip, but I haven't seen one.
londons_explore 3 hours ago [-]
The mel spectrum is the first part of a speech recognition pipeline...
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.
calibas 2 hours ago [-]
I was playing around with this recently, but the problem I encountered is that most AI analysis techniques like stem separation aren't built to work in real-time.
wolvoleo 1 hours ago [-]
Thanks for this! Exactly the thing I'm struggling with now. Making decent visualisation for music based on ESP32-S3.
panki27 3 hours ago [-]
Had a similar setup based on an Arduino, 3 hardware filters (highs/mids/lows) for audio and a serial connection. Serial was used to read the MIDI clock from a DJ software.
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.
More than 20 years ago or so I made a small LED display that used a series of LM567 (frequency detection ICs) and LM3914 (bar chart drivers) to make a simple histogram for music.
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.
8cvor6j844qw_d6 3 hours ago [-]
Are these available commercially for consumers?
p0w3n3d 4 hours ago [-]
IANAE but I would go for electric circuit, not electronic software that steers the led. I think that nowadays, with the LLM support it can be easier and better to optimise it for the sake of latency.
mrob 4 hours ago [-]
If you want minimum latency, you want the input side of an traditional vocoder, not an FFT. This is the part that splits the modulator signal into frequency bands and puts each one through an envelope follower. Instead of using the outputs of the envelope followers to modulate the equivalent frequency bands of a carrier signal, you can use them to drive the visualizer circuit.
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.
avisser 2 hours ago [-]
For recorded music, you could always buffer however many milliseconds of audio to account for the processing.
IshKebab 36 minutes ago [-]
It's not that hard. I did a real-time version of the Beatroot algorithm decades ago that worked pretty well for being such a simple algorithm.
askl 5 hours ago [-]
Interesting. I'm currently in the process of building something with a audio reactive LED strip but didn't come across this project yet.
The WLED [1] ESP32 firmware seems to be able to do something similar or potentially more though.
Edit: Oh wait, that project needs a PC or Raspberry PI for audio processing. WLED does everything on the ESP32.
stavros 4 hours ago [-]
Yeah WLED does it fine, I've built a few and it works well.
MrBuddyCasino 2 hours ago [-]
WLED is decent but tbh the lag is very noticeable. Did you compare to this python thing?
askl 52 minutes ago [-]
No, haven't tried it.
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)
turbine401 4 hours ago [-]
[dead]
mockbolt 2 hours ago [-]
[flagged]
isoprophlex 2 hours ago [-]
Are you using multiple accounts to post the same comment?!
kbouck 2 hours ago [-]
[flagged]
m3kw9 2 hours ago [-]
how is it hard, do a A to D, add a filter, do compute, then do D to A.
kennywinker 2 hours ago [-]
Not hard to do, hard to do well. Hiding all complexity with a hand wavey “do compute” doesn’t make that bit easy
m3kw9 2 hours ago [-]
Yeah i get it, the details are hard.
cogman10 2 hours ago [-]
The article covers that.
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.
Rendered at 16:16:31 GMT+0000 (Coordinated Universal Time) with Vercel.
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.)
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.
- 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.
[1]: https://www.pjrc.com/
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.
And of course, by the time I got it to work perfectly I never looked at it again. As is tradition.
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.
(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.
[1] https://cdn.sparkfun.com/assets/d/4/6/0/c/MSGEQ7.pdf
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.
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.
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.
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.
[1] https://kno.wled.ge/
Edit: Oh wait, that project needs a PC or Raspberry PI for audio processing. WLED does everything on the ESP32.
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)
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.