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!
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)
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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)
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.
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.
Another related project that builds on a similar foundation: https://github.com/ledfx/ledfx
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.
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.