Fun fact: I’m working on turning ESPectre into a Wi‑Fi Theremin (the musical instrument you play by moving your hands near an antenna).
The idea of “playing” by simply moving around a room sounds a bit ridiculous… but also kind of fun.
The key is the Moving Variance of the spatial turbulence: this value is continuous and stable, making it perfect for mapping directly to pitch/frequency, just like the original Theremin. Other features can be mapped to volume and timbre.
It’s pure signal processing, running entirely on the ESP32. Has anyone here experimented with audio synthesis or sonification using real-time signal processing?
Ive worked on some sonification projects that used signals from xbox kinect lidar, piezos, and other sensors. Co-author on paper i wrote developed a "strummable" theremin that divided physical space with invisible "strings" of various tunings. We preferred running synthesis on PC when possible and just outputting midi and OSC, as DSP on ESP32 has limits for what can be achieved in under 5-10ms. If the goal is hardware audio output, you may need to look into dedicated DSP chips and an audio shield for better DAC- but for prototyping can easily bang a square wave through any of esp32 pins
I wonder if somebody could make an open hardware version of the Leap Motion with this technique (though I'm not sure how accurate/repeatable the sensing is - Leap Motion could detect with an accuracy of < 0.7mm)
You hit the nail on the head! That's precisely the motivation.
Having two kids myself, I've thought of turning it into a game: blindfolded hide-and-seek where the pitch of the Wi-Fi Theremin tells the seeker how close they are to the 'signal disruption' of the other person. It's essentially a real-time sonar game!
Been working on this very idea casually for couple years with ESP-IDF and I could never get the statistical signal processing just right (by my definition). Things I've tried: adapting filtering (LMS, Kalman), kernel methods (NEWMA, MMD), detectors (CUSUM, GLR), dimensionality reduction (random projection, online PCA), whitening, etc.
I use a single ESP32 in STA/AP mode which sniffs ACK packets with a specific destination mac, which come from any server on my WiFi network (uses a special sniffing mode IIRC). This way I can receive regular CSI packets originating from a fixed location and doesn't need another device running.
I'll have to look at this code, maybe I just overlooked the obvious or my requirements were too high!
ESPectre takes a different architectural approach that might address some of the challenges you encountered:
1. Instead of STA/AP mode on a single ESP32, ESPectre uses the natural traffic between your existing router and an ESP32-S3 in station mode. To ensure a stable, continuous CSI packet rate, I implemented a traffic generator that sends ICMP pings to the gateway at a configurable rate (default: 20 pps). This provides bidirectional traffic (request + reply) that reliably triggers CSI generation, giving you predictable packet timing without relying on ambient network traffic or special sniffing modes.
2. Rather than applying filters directly to raw CSI, ESPectre uses Moving Variance Segmentation (MVS) on unfiltered spatial turbulence (std dev of subcarrier amplitudes).
3. The filters are applied to features, not to the segmentation signal itself. This preserves motion sensitivity while cleaning up the feature data
I found that having a stable transmitter (the router) combined with controlled traffic generation provides more consistent multipath patterns and predictable CSI timing, which makes the segmentation more reliable.
Actually I misspoke. I previously used STA/AP mode (and two ESP32s) but I switched to something close to what you describe. I filter the pings to only get the ones targeting a specific MAC (in promiscuous mode). This way I get only specific CSI packets and they're perfectly periodic at whatever rate I want.
Sounds like your MVS approach is a sliding window variance of the cross channel variance, with some adaptive thresholding. My pre-processing has generally been an EWMA de-meaning filter followed by some type of dimensionality reduction and feature extraction (kernel or hand-crafted, like raw moments), which I think fits into your overall architecture.
I'll have to look more closely at your work, thanks for sharing!
Interesting note, I actually disabled promiscuous mode after some testing because it made the CSI signal noisier and consumed more resources.
I found that normal station mode with pings to gateway gave me cleaner, more predictable CSI data. But your MAC filtering approach might mitigate those issues!
You're spot on about the MVS approach. It's essentially a sliding window variance of the spatial turbulence (std dev across subcarriers), with adaptive thresholding based on the moving variance of that signal.
If you're interested in the MVS details, I wrote a free Medium article that walks through the segmentation algorithm step-by-step with visualizations. Links are in the README.
Your approach is actually quite similar to what I'm doing, just in a different order:
- My flow: Raw CSI → Segmentation (MVS) → Filters (Butterworth/Wavelet/Hampel/SG) → Feature extraction
- Your flow: Raw CSI → EWMA de-meaning → Dimensionality reduction → Feature extraction
The main difference is that I segment first to separate IDLE from MOTION states (keeping segmentation on raw, unfiltered CSI to preserve motion sensitivity), then only extract features during MOTION (to save CPU cycles).
Thanks for the thoughtful feedback! Always great to exchange notes with someone who's been in the trenches with CSI signal processing
The final statistic I threshold is the Moving Variance of Spatial Turbulence.
The decision is a binary comparison: When moving_variance > threshold then MOTION state (movement detected) else IDLE state.
The features are extracted only during MOTION segments (to save CPU cycles) and published via MQTT.
They serve as rich foundation data for potential external ML models (e.g., to capture nuances like gestures, running, or falling), but they are absolutely not used for the core segmentation decision itself.
This jogged some memories! Was working on something that used radio waves to detect objects and humans ~5 years ago, I see that we've come a long way since then.
One of our goals(abandoned) was to also extend to wifi routers, so I am excited to see continued interest in this space!
I see you have done a fair amount of work to document calibration for various scenarios. Have you tried to calibrate to ignore cats? Can you prioritize different algorithms to focus on size rather than speed of movement?
Also, I use an ebay purchased ruckus router designed for commercial settings. Will the stronger signal and beam forming from the router provide better or worse performance, or is that mainly down to the esp32?
Tommysense creates a sensing mesh between devices, while ESPectre uses your existing Wi-Fi router as the transmitter. As a result, ESPectre needs only one device per area but requires a compatible router with solid 2.4 GHz coverage. The overall goal is similar, but ESPectre is open-source!
It's neat that Tommysense works on top of esphome... I'm currently using Bermuda BLE trilateration, but it doesn't quite work, especially in a multy-story living space (e.g. a townhouse). So I already have a bunch of esphome Bluetooth proxies all over the building.
But no source and "lifetime license if you join our discord" is kinda not my jam.
You need one sensor for each area you want to monitor independently.
With devices more capable than the ESP32‑S3, the coverage would likely be greater.
It really depends on the environment: wall materials, interference, signal strength, and even temperature and humidity all play a role. Honestly, 100 m² is probably too large for a single sensor to cover reliably.
That said, ESP32 boards are very inexpensive, you can find them online for around €1 or even less.
Is the ESP-S3 your recommendation? Or should one look into the ESP-C6 / another model? The prices I'm seeing here in Sweden are more around €10 rather than €1, but I guess that's still much less than the average presence detector.
The price difference can vary wildly by region, but €10 is still super cheap compared to a dedicated presence sensor!
Regarding the chip recommendation, it’s a classic trade-off:
ESP32-S3: This is what I mainly recommend today. It gives you more raw processing power (dual-core), which is great for managing the OS, MQTT, and the CSI analysis without hiccups.
ESP32-C6: I just ordered a C6 myself to run extensive tests! The C6 might have a superior quality Wi-Fi module and better internal CSI extraction capabilities due to hardware improvements.
Using two ESP32-S3 modules you can get ~6000 packets per second with CSI data. I'm using this as a cheap replacement for specialty high-G gyroscope modules, but it could see use for this type of motion detection as well.
The project’s open-source nature acts as an ethical safeguard, and I am explicitly not pursuing any identity recognition features, just movement detection.
But you are absolutely right that, in theory, misuse of this technology could reveal certain behavioral patterns that might lead to identification.
However, it can also be extremely useful for safety purposes, for example, detecting people during a house fire or an earthquake.
Sure, the ESP32 will connect to whichever mesh node provides the best 2.4 GHz signal
- It monitors CSI from that specific node (the one it's associated with)
- If the ESP32 roams to a different mesh node, it will start monitoring CSI from the new node
The system doesn't care about the router's internal mesh topology, it just needs a stable connection to receive CSI data from the associated access point.
Fun fact: I’m working on turning ESPectre into a Wi‑Fi Theremin (the musical instrument you play by moving your hands near an antenna).
The idea of “playing” by simply moving around a room sounds a bit ridiculous… but also kind of fun.
The key is the Moving Variance of the spatial turbulence: this value is continuous and stable, making it perfect for mapping directly to pitch/frequency, just like the original Theremin. Other features can be mapped to volume and timbre.
It’s pure signal processing, running entirely on the ESP32. Has anyone here experimented with audio synthesis or sonification using real-time signal processing?
Ive worked on some sonification projects that used signals from xbox kinect lidar, piezos, and other sensors. Co-author on paper i wrote developed a "strummable" theremin that divided physical space with invisible "strings" of various tunings. We preferred running synthesis on PC when possible and just outputting midi and OSC, as DSP on ESP32 has limits for what can be achieved in under 5-10ms. If the goal is hardware audio output, you may need to look into dedicated DSP chips and an audio shield for better DAC- but for prototyping can easily bang a square wave through any of esp32 pins
I wonder if somebody could make an open hardware version of the Leap Motion with this technique (though I'm not sure how accurate/repeatable the sensing is - Leap Motion could detect with an accuracy of < 0.7mm)
I'm sure the kids will love this! Wi-Fi Theremin sounds great.
You hit the nail on the head! That's precisely the motivation.
Having two kids myself, I've thought of turning it into a game: blindfolded hide-and-seek where the pitch of the Wi-Fi Theremin tells the seeker how close they are to the 'signal disruption' of the other person. It's essentially a real-time sonar game!
Been working on this very idea casually for couple years with ESP-IDF and I could never get the statistical signal processing just right (by my definition). Things I've tried: adapting filtering (LMS, Kalman), kernel methods (NEWMA, MMD), detectors (CUSUM, GLR), dimensionality reduction (random projection, online PCA), whitening, etc.
I use a single ESP32 in STA/AP mode which sniffs ACK packets with a specific destination mac, which come from any server on my WiFi network (uses a special sniffing mode IIRC). This way I can receive regular CSI packets originating from a fixed location and doesn't need another device running.
I'll have to look at this code, maybe I just overlooked the obvious or my requirements were too high!
ESPectre takes a different architectural approach that might address some of the challenges you encountered:
1. Instead of STA/AP mode on a single ESP32, ESPectre uses the natural traffic between your existing router and an ESP32-S3 in station mode. To ensure a stable, continuous CSI packet rate, I implemented a traffic generator that sends ICMP pings to the gateway at a configurable rate (default: 20 pps). This provides bidirectional traffic (request + reply) that reliably triggers CSI generation, giving you predictable packet timing without relying on ambient network traffic or special sniffing modes.
2. Rather than applying filters directly to raw CSI, ESPectre uses Moving Variance Segmentation (MVS) on unfiltered spatial turbulence (std dev of subcarrier amplitudes).
3. The filters are applied to features, not to the segmentation signal itself. This preserves motion sensitivity while cleaning up the feature data
I found that having a stable transmitter (the router) combined with controlled traffic generation provides more consistent multipath patterns and predictable CSI timing, which makes the segmentation more reliable.
Actually I misspoke. I previously used STA/AP mode (and two ESP32s) but I switched to something close to what you describe. I filter the pings to only get the ones targeting a specific MAC (in promiscuous mode). This way I get only specific CSI packets and they're perfectly periodic at whatever rate I want.
Sounds like your MVS approach is a sliding window variance of the cross channel variance, with some adaptive thresholding. My pre-processing has generally been an EWMA de-meaning filter followed by some type of dimensionality reduction and feature extraction (kernel or hand-crafted, like raw moments), which I think fits into your overall architecture.
I'll have to look more closely at your work, thanks for sharing!
Interesting note, I actually disabled promiscuous mode after some testing because it made the CSI signal noisier and consumed more resources. I found that normal station mode with pings to gateway gave me cleaner, more predictable CSI data. But your MAC filtering approach might mitigate those issues!
You're spot on about the MVS approach. It's essentially a sliding window variance of the spatial turbulence (std dev across subcarriers), with adaptive thresholding based on the moving variance of that signal.
If you're interested in the MVS details, I wrote a free Medium article that walks through the segmentation algorithm step-by-step with visualizations. Links are in the README.
Your approach is actually quite similar to what I'm doing, just in a different order:
- My flow: Raw CSI → Segmentation (MVS) → Filters (Butterworth/Wavelet/Hampel/SG) → Feature extraction
- Your flow: Raw CSI → EWMA de-meaning → Dimensionality reduction → Feature extraction
The main difference is that I segment first to separate IDLE from MOTION states (keeping segmentation on raw, unfiltered CSI to preserve motion sensitivity), then only extract features during MOTION (to save CPU cycles).
Thanks for the thoughtful feedback! Always great to exchange notes with someone who's been in the trenches with CSI signal processing
I noticed your feature vector is large and you don't use ML. What's the final statistic that you threshold?
The final statistic I threshold is the Moving Variance of Spatial Turbulence.
The decision is a binary comparison: When moving_variance > threshold then MOTION state (movement detected) else IDLE state.
The features are extracted only during MOTION segments (to save CPU cycles) and published via MQTT.
They serve as rich foundation data for potential external ML models (e.g., to capture nuances like gestures, running, or falling), but they are absolutely not used for the core segmentation decision itself.
This jogged some memories! Was working on something that used radio waves to detect objects and humans ~5 years ago, I see that we've come a long way since then.
One of our goals(abandoned) was to also extend to wifi routers, so I am excited to see continued interest in this space!
https://www.sensorsportal.com/HTML/ST_JOURNAL/PDF_Files/P_32...
I see you have done a fair amount of work to document calibration for various scenarios. Have you tried to calibrate to ignore cats? Can you prioritize different algorithms to focus on size rather than speed of movement?
Also, I use an ebay purchased ruckus router designed for commercial settings. Will the stronger signal and beam forming from the router provide better or worse performance, or is that mainly down to the esp32?
> The system does NOT use Machine Learning, it relies purely on Math.
You may be surprised to find out how machine learning works!
That's a fair point, and as a math graduate, I absolutely agree that ML is fundamentally applied math.
When I say 'No ML,' I mean there is no training phase, no labeled data needed, and no neural network model used to infer the rules.
The distinction here is that all the logic is based purely on signal processing algorithms.
Thanks for raising the point!
Interesting! Are you familiar with tommysense.com? I think it doing something similar? Did not yet have time to try it.
Tommysense creates a sensing mesh between devices, while ESPectre uses your existing Wi-Fi router as the transmitter. As a result, ESPectre needs only one device per area but requires a compatible router with solid 2.4 GHz coverage. The overall goal is similar, but ESPectre is open-source!
It's neat that Tommysense works on top of esphome... I'm currently using Bermuda BLE trilateration, but it doesn't quite work, especially in a multy-story living space (e.g. a townhouse). So I already have a bunch of esphome Bluetooth proxies all over the building.
But no source and "lifetime license if you join our discord" is kinda not my jam.
Amazing stuff!
Am I right in understanding that only a single ESP32 device is needed (plus a router)?
Probably one per room because all the examples at https://github.com/francescopace/espectre/blob/main/CALIBRAT... are about a single room.
Is the author reads this, how does the system cope with multiple rooms in the same house, maybe a two or three storeys house?
Yes, you’re both correct:
You need one sensor for each area you want to monitor independently. With devices more capable than the ESP32‑S3, the coverage would likely be greater.
The ESP32‑C6, in particular, offers significantly better performance. Check out this comparison video from Espressif: https://www.youtube.com/watch?v=JjdpzM6zVJ8
How about if I want to monitor the apartment as a whole? Would 100m2 be too much for one sensor to handle?
It really depends on the environment: wall materials, interference, signal strength, and even temperature and humidity all play a role. Honestly, 100 m² is probably too large for a single sensor to cover reliably.
That said, ESP32 boards are very inexpensive, you can find them online for around €1 or even less.
Is the ESP-S3 your recommendation? Or should one look into the ESP-C6 / another model? The prices I'm seeing here in Sweden are more around €10 rather than €1, but I guess that's still much less than the average presence detector.
The price difference can vary wildly by region, but €10 is still super cheap compared to a dedicated presence sensor!
Regarding the chip recommendation, it’s a classic trade-off:
ESP32-S3: This is what I mainly recommend today. It gives you more raw processing power (dual-core), which is great for managing the OS, MQTT, and the CSI analysis without hiccups.
ESP32-C6: I just ordered a C6 myself to run extensive tests! The C6 might have a superior quality Wi-Fi module and better internal CSI extraction capabilities due to hardware improvements.
Using two ESP32-S3 modules you can get ~6000 packets per second with CSI data. I'm using this as a cheap replacement for specialty high-G gyroscope modules, but it could see use for this type of motion detection as well.
Are you using two separate S3s as a dedicated Transmitter/Receiver pair, or are both transmitting data simultaneously?
The surveillance implications for this technology are fascinating and frightening.
The project’s open-source nature acts as an ethical safeguard, and I am explicitly not pursuing any identity recognition features, just movement detection.
But you are absolutely right that, in theory, misuse of this technology could reveal certain behavioral patterns that might lead to identification.
However, it can also be extremely useful for safety purposes, for example, detecting people during a house fire or an earthquake.
Would this work with a mesh router?
Sure, the ESP32 will connect to whichever mesh node provides the best 2.4 GHz signal
- It monitors CSI from that specific node (the one it's associated with)
- If the ESP32 roams to a different mesh node, it will start monitoring CSI from the new node
The system doesn't care about the router's internal mesh topology, it just needs a stable connection to receive CSI data from the associated access point.
my brain expected the part after `-` to be new speculative execution bugs
Hahaha, I love your brain!