Application of neural networks in DIY

It is not surprising that AI functionality is now available in DIY devices. Let's see what can be purchased from such devices right now.

Gravity: Offline Voice Command Recognition Module

Link: https://www.dfrobot.com/product-2665.html

Approximate price: 17$

The voice command recognition module works completely offline, equipped with a built-in speaker and microphone, and allows you to add your own commands.

Compatible with Arduino, micro:bit, and ESP32. Interesting for use in "smart" homes, voice assistants, and generally for any other interaction between humans and devices.

In addition to 121 basic commands in English such as "turn on the light", "open the door", "reduce the volume", etc. (the full list of commands can be found here), it is possible to add 17 new ones in any language. As an option, you can even record a sound: a clap of hands, a click, or a cat's meow (for example, to activate the feeder, if, of course, you can get the pet to meow three times for training).

A bonus is the module's real-time voice notification of whether the command was accepted and understood.

Among the advantages that caught my eye personally:

• changing the wake-up word – definitely a turnkey setting;

• detailed documentation and wiki;

• variability of use and level of individualization of the solution – in the house and beyond, you can make any part of the household "smart" if there is a technical possibility.

Cons: lack of Russian language.

Technical specifications

• Operating voltage: 3.3 – 5V

• Maximum operating current: ≤370 mA (5V)

• Interfaces: I2C/UART

• I2C address: 0x64

• Microphone sensitivity on board: -28 dB

• Dimensions: 49×32 mm / 1.93×1.26”

• Operating temperature: 0-70℃

Commands:

• Fixed: 121

• Fixed activation word: 1

• Customizable commands: 17

• Command to activate training: 1

To be fair, it should be noted that 9 commands used in the training process are not included here.

AI module: Voice control of the robot

Link: https://ai-mcu-hub.ru/product/1584421305/

Price: 1$6.06.

This product partly overlaps with the previous one, but it is already tailored for a specific task - robot control (although it can be adapted for anything). The solution is implemented on the ESP32 board, supports 9 commands (mainly direction of movement and speed changes: "forward", "backward", "start", "stop", "slow" etc.), and a word to go to sleep mode. The board has a built-in microphone, an amplifier for connecting a speaker, and LEDs, which, by the way, are very useful, but more on that later.

It seems to me that this product is perfect for schoolchildren and novice circuit designers. The functionality, of course, is much simpler than that of the first "lot", but the clarity is much higher.

Firstly, the system's response not only through the speaker but also through the LEDs is a great idea. The speaker can be quiet, and in case of a breakdown, feedback with the device is completely lost. Here everything is extremely clear: green - accepted-understood, red - does not understand, needs to be repeated. Perhaps the first product also has LEDs, correct me if I'm wrong, but I couldn't find them. Secondly, the speaker here says not just "Ok", but pronounces the commands, which is also a plus for clarity.

I like the ease of implementation of the idea. In fact, any children's robotics kit is enough to start working.

Technical specifications

• Microcontroller: ESP32-S3;

• Memory: 384 KB ROM and 512 KB SRAM

• MEMS microphone: MP34DT06JTR

• LEDs

• Sound I2S amplifier D-class on MAX98357

• Interfaces: UART, SPI, I2C

• Power supply: 5V.

• Average power consumption - 30 mA

Interactive game "Teacher"

Link: https://ai-mcu-hub.ru/product/1690180792/

Continuing the hit parade of AI devices for word recognition.

Another solution that is tailored to a specific task - learning English at the elementary level. In fact, this is not quite a DIY device, but a full-fledged solution that, perhaps, just needs to organize word package updates. At least I don't quite understand where it could be integrated, unlike previous modules, but the idea itself is so simple and attractive that it seemed a sin to leave this option out of the review.

So, what do we have. A board with a module for recognizing a set of English words. There are 40 words in the package, 10 of which are numbers, but the manufacturer promises to expand this package in the future. Built-in speaker that: a) gives feedback on attempts to pronounce the word, b) a hint, c) a sound association. The monitor displays pictures or examples that need to be recognized or counted and pronounced in English. And again we have LEDs, which also make it easy to understand whether the attempt was successful. Powered by a power bank (type C) or via USB. The essence of the game: to pronounce the desired word in English without distorting the phonetics. The module, judging by the video, is not too picky about pronunciation, but it does not allow outright "Runglish".

I wish I could get my hands on it to play... although I'm not a child and my level is still above elementary.

Technical specifications

(although they don't matter much here)

• Microcontroller: ESP32 S3 and ESP32 C3

• MEMS microphone: MP34DT06JTR

• LEDs: RGB LED - SK6805

• Power supply: USB Type-C with smartphone adapter or computer USB port. 5 V, 500 mA (min).

• 4 buttons: Reset, Back and 2 User/Boot

• OLED display

Gravity: Huskylens - AI camera

Link: https://www.dfrobot.com/product-1922.html

Approximate price: 55$

We move from speech and sound recognition to image and object recognition. And here is a kind of "Swiss knife" in image recognition. The module based on Kendryte K210 offers a bouquet of AI capabilities: object tracking, face recognition, object recognition, line tracking, color and tag recognition. Compatibility with Arduino, micro:bit, Raspberry Pi and LattePanda is declared, and, as it seems to me, it opens up huge opportunities for creating your own solutions. The module can be trained for your requests without connecting to the Internet.

The instructions state that the module can be trained to recognize certain gestures, which, in addition to the banal use in "smart" homes, can be a breakthrough for people with disabilities (mute or deaf). Line or markup tracking and color recognition is a great option for developing a variety of road solutions. There are millions of applications.

That is why this product received the title of "Swiss knife" from me. If it could also embroider on a sewing machine, it would be priceless! Although, if you drive the line recognition algorithm to read knitting patterns..., then Matroskin may have to make room on his pedestal.

Technical specifications

• Processor: Kendryte K210

• Image sensor: OV2640 (2.0 megapixel camera)

• Operating voltage: 3.3~5.0 V

• Current consumption: [email protected], [email protected] (in face recognition mode; 80% backlight brightness; light off)

• Interfaces: UART, I2C

• Display: 2-inch IPS screen with a resolution of 320*240

• Built-in algorithms: Face recognition, Object tracking, Object recognition, Line tracking, Color recognition, Tag recognition

• Dimensions: 52mm * 44.5mm / 2.05 x 1.75 inches

In conclusion

Of course, four samples are far from everything that the market can offer to those who are looking, but in my opinion, these are a good start and open up rich opportunities for the introduction of AI into our lives. And what products caught your attention? I would be grateful if you share your findings in the comments.