Have you ever always wanted a very very small voice note recorder that would fit in your pocket? Something that would always work, and always be available to take a note at the touch of a button, with no fuss? Me neither.

Until, that is, I saw the Pebble Index 01, then I absolutely needed it right away and had to have it in my life immediately, but alas, it is not available, plus it’s disposable, and I don’t like creating e-waste. What was a poor maker like me supposed to do when struck down so cruelly by the vicissitudes of fate?

There was only one thing I could do:

I could build my own, shitty version of it for $8, and that’s exactly what I did.

The problem

Like everyone else, I have some sort of undiagnosed ADHD, which manifests itself as my brain itching for a specific task, and the itch becoming unbearable unless I scratch it. This usually results in me getting my phone out, no matter where I am or who I’m with, and either noting stuff down or doing the task, which some people perceive as rude, for inexplicable reasons that are almost certainly their fault. Because, however, it has proved easier to just not get my phone out in polite company than convince everyone of how wrong they are, I just do the former now, but that makes the itch remain.

Also, sometimes I’m just in the middle of something, and an idea pops into my head for later pursuit, but I get distracted by a squirrel, a car going by, or the disturbing trend of the constant and persistent erosion of civil rights all over the world, and I forget the idea.

The Pebble Index showed me that there’s a better way, a device that’s unobtrusive, available, and reliable enough that I could just press a button, speak into it, and know for sure that my sonorous voice would reach the bowels of my phone, where it would be stored safely until I was bored and wanted something to do.

I didn’t want to have to get my phone out, unlock it, open a voice recorder app, hold down a button, speak, wonder if it heard me, look at the button, realize I had already pressed it, press it again, say the thing again, press it again to stop, exit the app, lock my phone, and put it back into my pocket.

I wanted to take a thing out, press a button, speak, release the button, done.

The solution

The initial thinking was that I’d use a microcontroller (an ESP32 is my microcontroller of choice these days), a microphone, and a lithium battery, and that’s basically all the hardware this needs! Most of the heavy lifting would need to be done in software. This would need:

• A way for the device to record audio onto some sort of persistent storage, for the case where you didn’t have your phone close to you.
• A way for the device to sleep, consuming almost no power, until it was woken up by the button.
• A way to transfer the files from the device to the phone, for later listening.
• A battery indicator would be very nice, so I knew when to recharge it.

Luckily, I know enough about electronics to know that LLMs would definitely know how to build something like that.

The hardware

Indeed, Claude confirmed my suspicions by saying that all I need is a microphone and an ESP32. It recommended an ESP32-C6 but I went with an ESP32-S3, as it had an onboard charge controller and would be able to charge a lithium battery from USB, which is very handy when you’re making a thing that runs on battery. The ESP32 is a microcontroller, a little computer that’s just really small. The main difference of the S3 from the C6 is that the S3 is more capable, and has more power.

I keep an assortment of random components around, so I had an ESP32-S3 board. It’s a no-name, crappy one from AliExpress, not a good, Seeed-branded one from AliExpress, but it would have to do.

Unfortunately, I didn’t have a MEMS microphone (which is basically an angelic grain of rice that can hear, with excellent quality), but I did have an electret mic, which is huge and bad quality and would sound like an old-timey radio, but it was there and it was ready and it was willing, and after a few beers it seemed like it was right, or at least right for right now.

I also had a very thin LiPo battery, which would suit very well. For the final device I’d want a battery that’s a tiny bit shorter, as this one was around 40% longer than the ESP32, but it would do great for now.

Hooking the stuff together

I quickly soldered everything together and recorded some audio. It worked! It worked and nobody was going to take that from me, even though it was crackly and the quality wasn’t great.

Unfortunately, at this stage I realized that the analog electret microphone consumes too much energy, even when sleeping, which is terrible on a device that would spend more time sleeping than the beauty from that fairytale, Sleepy the Dwarf. To counteract that, I decided to use a MOSFET to cut power to the mic when the device was asleep. A MOSFET is a little switch that you can turn on and off from a microcontroller, basically.

Full disclosure here, before using the MOSFET to turn the mic on and off, I went down a multi-hour rabbit hole trying to design a latching circuit that would allow the ESP32 to turn itself off and consume almost no power. Instead, it consumed a lot of my time, without anything to show for it, because I didn’t manage to make it work at all.

The MOSFET for the mic worked fairly well, though, and the device didn’t consume much power when asleep. The real gains, however, were going to be had when the MEMS microphone I ordered arrived, as those use infinitesimal amounts of current when asleep, and have much better sound quality as well, as they are digital.

Problems

The analog microphone crackled and popped and took a while to stabilize after boot, which was unfortunate because I wanted the device to be ready as soon as the user pressed the button. There was also a recording bug where the recording was missing a few milliseconds of audio every so often, which led to dropped phonemes and words sometimes sounding like other words because parts of them were dropped.

All these problems were weird enough and hard enough to debug that I resolved to just wait for my digital MEMS microphone to arrive, which would solve them in one fell swoop, as it is digital and amazing.

The curious case of the listening device

After the relatively easy part of connecting a few wires together, now came the hard part: Designing a case for the whole thing that would fit without leaving much empty space, to make the device as small as possible. This was very hard to do with this massive microphone that was as tall as everything else (including battery) combined.

I initially tried to point the microphone downward while mounting it at the top, so it would take up the least amount of vertical space possible, but the PCB made that hard, as the microphone was soldered to it. I ended up desoldering the mic from the PCB, trimming the PCB to make it shorter, and connecting the mic to it with wires. That allowed me to make the case (and thus the device) smaller, but at what cost? Nothing, turns out, because it worked great.

The Bluetooth script

The device was working great, but I didn’t want it tethered to my computer, I wanted to be able to take it out and about and show it the wonders of the world. To do this, I needed Bluetooth. Unfortunately, I have exactly zero idea how Bluetooth works, and would need to spend days or weeks figuring stuff out, but, luckily for me, I had a Claude subscription.

It took a bit of back-and-forth, but I did manage to end up with a Python script that would connect to the pendant, download the audio files, and convert them from ADPCM to MP3, for expanded compatibility. To maximize battery life, the way things worked was:

• You pressed the button.
• If you held it down for more than half a second, the recording would “count”.
• If there was a recording made (i.e. if you held the button down long enough), it would be saved.
• Bluetooth would turn on and look for a phone or computer that’s ready to receive.
• The device would send the file and go to sleep again.

This worked really well, the device was awake for a small amount of time (10 seconds), but it could be awoken at any time just by tapping the button. At that point, it would transfer to the PC any files that were on the pendant, and go back to sleep.

One downside was that transfers would take an inordinate amount of time, sometimes reaching 2 minutes for a 10-second clip. OpenAI’s Codex was really helpful here, finding a solution for fast BLE transfers that made sending files 100x faster than it was before.

The new microphone

Because I’m too impatient to wait for the slow boat from China, I ordered the same microphone locally. I had to pay an arm and a leg in shipping and impatience fees, but it was worth it, because I finally had a MEMS mic! It’s so cute and tiny, I immediately found a spot for it over the board, added the switch, added a voltage divider for sensing battery voltage, and that was it!

The new mic sounds fantastic, it sounds better than recording with your phone, for some odd reason that I’m sure is all in my head. What’s more, it doesn’t have the weird bugs that plagued me with the analog mic.

The second-gen case

With this smaller mic, I could now design a better case. I designed the case you see on the right, which is the second generation. There will be a third, when I receive the shorter battery, which means I will have a choice of either making the device longer but half as thick, or around 40% shorter. I think I will go for longer but thinner, I’d quite prefer to have a thin device in my pocket, even if it’s long, than a stubby one that pokes out.

Still, the new battery (and the new case) will mark the completion of this project and make me a very happy man. For the second-gen case, I decided to jazz it up and add a red stripe around it, because it was easy to do and because I think it looks good. Unfortunately, the feature I wanted most (fillets, i.e. rounded corners) wasn’t possible due to the lack of empty space inside the case. I hope the final device will have some more space for fillets, at least.

The Android app

Once I was done with the device, it was time to make it more ergonomic: I’d need to create an Android app so I wouldn’t have to wait to get to my PC. I also knew I wanted note transcription, as it’s really useful to be able to see what you said without having to listen to the audio again.

Unfortunately again, I have no idea about Android development, only having written a small app years ago. Fortunately, though, Claude turned out to be pretty good at it, and one-shotted this app that you see here.

For the transcription, I used GPT-4o Transcribe, which is great and understands both English and Greek, languages I fail to speak in equal measure. I have to say, it’s pretty magical to speak into a little box and to see the audio already captured and transcribed on your phone.

With the Android app, I could now test the device in real-world use. One thing I noticed is that battery dies way too fast. I suspect that has something to do with the cheap board, so I’ve ordered an original Seeed Xiao board, and I hope that will fix the problem once and for all, as they advertise low power usage and they’re a trustworthy brand.

I also added a “webhook” convenience function to the Android app, so that the latter would be able to send the transcription to a server for further processing.

The device is extremely reliable, which makes me a lot more likely to use it. I know that, if I press the button, the audio will be recorded and stored, and nothing will happen to it, which makes for a very relaxed and calming experience.

Before I continue, I want to say you can find all the files in this project (firmware, Android app, whatever else) in its GitHub repository:

https://github.com/skorokithakis/middle

That’s right, I called it Middle, because it was the next thing after the Index. I know it’s a silly name, I don’t care, don’t use it, I’m not changing it.

The AI agent

In the “draw the rest of the fucking owl” portion of this article, I realized I didn’t want the notes to just go to my phone when LLMs exist. I wanted an LLM to take the notes and do something with them, so I spent a few weeks writing an AI agent that’s more useful than what currently exists. The device’s Android app sends the transcribed text to this AI, which processes it.

I’m going to write another post about this, but basically, I wanted an AI personal assistant that could help with all the little chores in my life. AI assistants are interesting because they’re:

• Very open-ended tools, and
• Highly personal.

This means that, when everyone inevitably asks “what is it good for”, I can’t really give a good answer, because the answer is “it takes care of all the little annoyances for me”, but nobody has the same annoyances and can’t really imagine what the bot does, so they don’t engage with it.

The amazing thing for AI assistants for me is the fact that they can string together multiple (otherwise small) tools to do something that’s more valuable than the sum of its parts. For example, I asked the agent to give me a daily briefing every morning, consisting of my todos for the day, my calendar events, whether any refund has hit my bank, and whether any packages are due to be delivered today.

The agent also checks my gym bookings and asks me every morning if I do plan to go, or if I intend to cancel. If I tell it to cancel, it does, but if I say I’ll go, it sets an alarm for a few minutes before, which I’m much more likely to see than my calendar’s one. It will also (entirely of its own volition) mention things like “you have a gym booking today 7-8pm but you have a restaurant booking at 9pm and it’ll take you more than an hour to shower and make it”, which a regular calendar wouldn’t be able to figure out.

I’ve made it fantastically secure, everything is sandboxed and you can run it on your laptop without fear. I use it constantly throughout the day for many little things, and the integration with the device takes the whole setup to another level.

You can find the bot here:

https://github.com/skorokithakis/stavrobot

Do let me know if you try it, it’s like OpenClaw but won’t steal your data and eat your firstborn.

Epilogue

If you have any ideas, feedback, flamebait, or whatever, you can Tweet or Bluesky me, or email me directly.