This is a proof of concept for my sister, she works as a teacher and while teaching her child to read had an idea, to build a device that would spell/vocalize the assembled syllables. So this is how I went about building the prototype to test the concept.
These components where used in the assembly:
- ESP32
- Audio board DY-SV5W
- 18650 li-ion battery (Reused Samsung 18650 from and old laptop battery)
- Kailh box red button for a mechanical keyboard
- Generic power switch
- 3D printed case
- No name random speaker ~90mm
- 1-cell USB-C charger board TP4056
- Boost/step-up converter MT3608
- 2x mini RC522 NFC readers
- 32x mini Ntag215 NFC cards one for each letter
ESP32 was chosen because it was available in the magical place that is my drawer. But seriously it’s like 1000x times faster than it needs to be, the main MCU does almost nothing but read button state, if pressed reads NFC tags then parses their data and instructs the audio board to play the correct file.
The audio board was chosen mainly for the low cost, and seemingly easy interface (as far as i could infer from the aliexpress listing). It’s fine even tho it has 2 draw backs:
- The command to get the current “play state” – if it is playing audio or not, doesn’t work and just interrupts the currently playing audio file. Didn’t test with longer audio files than a 2-3 seconds, but beware.
- It’s quite annoying to work with due to the filesystem used FAT32 and it not sorting the files. Meaning if you upload 10 files with name 1.mp3 … 10.mp3 and then add file 0.mp3 if you command file 1 to be played it will play the file 1.mp3 and to get the file 0.mp3 to be played you will need to ask for file 11. It acts this way because it doesn’t sort the files, so the order the files get copied onto the sd card matters. But using nautilus on ubuntu always uploaded in the order that the folder was sorted already so that wasn’t a big issue.
The selection for the switch is quite interesting, because all the switches i could get my hands on was quite stiff until I remembered that I have some Kailh box reds that I bought for testing a very soft, linear, smooth, long travel, mechanical switch that has a very mild water resistance due to its construction. It was perfect, even tho it looks bad, but looks don’t matter at this stage, because it’s very easy to press, and easy to see/understand what should be pressed.
The TP4056 charging board works fine, quite annoying to mount, because it has no mounting holes and the usb-c ports doesn’t protrude from the end of the PCB all that much, therefore the plastic is very thin around the port. And it gets HOT like really hot, like 20seconds of charging and you cannot hold your finger on the board for longer than 500ms hot. It’s fine I because everything is printed from PETG and and I charged it multiple times, but next time I would use something like this with lowered current and the heatsink.
RC522 worked great when I fixed the library that I found, because it had all kinds of nasty side affects to variables it was given and some retry logic. Range is good, they don’t interfere with each other even tho they are right next to each other.
The audio files where generated by supplying a TTS AI model with all valid combinations of 2 Lithuanian letters , and the baseline 32 sounds of single letters, the resulting list was around 650 sound files. But the AI generated lost of incorrect results, fails to generate or even straight up yells insults at you… Sooo, if this thing proves that it has any promise at all, the sound files would have to be regenerated, or just recorded professionally.
Leave a Reply