UNIHIKER K10 Powers Real-Time AI Sound Assistant
A maker project published on Hackster turns the UNIHIKER K10 learning board into a real-time AI sound assistant that recognizes household noises and shows them on screen, aimed at helping people with hearing loss. Built as a school project by a computer science teacher and two of his first-year middle-school students, the device listens continuously and, when it detects a sound like a doorbell, phone, intercom, or breaking glass, flashes its RGB LEDs and displays the event on its 2.8-inch screen. The team trained a TinyML classifier with Edge Impulse and runs inference locally on the board's ESP32-S3, triggering an alert when the model's confidence passes 75 percent. It also doubles as a basic security monitor by flagging sounds such as glass breaking. The write-up details practical fixes for audio capture, memory, and false positives.
The project
Published on Hackster, the UNIHIKER K10 AI Sound Assistant is a do-it-yourself system that uses the K10 learning board to recognize everyday household sounds and turn them into visual alerts. It was built as an end-of-year school project by a computer science teacher and two of his first-year middle-school students. The stated goal is accessibility: helping people with hearing loss notice sounds they might otherwise miss, such as a doorbell, a ringing phone, an intercom, or breaking glass. The board listens continuously and, when it recognizes a target sound, flashes its onboard RGB LEDs and shows the event on its screen.
How it works
The team trained a TinyML model with Edge Impulse, using Mel-Frequency Energy signal processing and a neural-network classifier, then exported it to run locally on the board with no cloud connection. In operation, the device stays in a low-power listening mode that watches ambient volume; once sound crosses a threshold above the measured background noise, it captures an audio buffer and runs on-device inference, triggering an alert only when the model's confidence exceeds 75 percent. The write-up documents real engineering hurdles and fixes, including moving audio buffers to PSRAM to avoid memory errors on the ESP32-S3, adding a dynamic sensitivity offset to cut false positives, matching the microphone's sampling rate to the model, and reducing screen flicker by redrawing only when the detected sound changes.
The hardware
The UNIHIKER K10, made by DFRobot, is an education-focused AI board built around an ESP32-S3 with Wi-Fi and Bluetooth, a 2.8-inch color screen, a camera, a MEMS microphone, a speaker, an RGB LED, and onboard temperature, humidity, light, and motion sensors. It can be programmed with the Scratch-based UNIHIKER MindPlus, the Arduino IDE, or MicroPython, which is part of why the teacher frames it as useful across beginner-to-advanced computer science lessons.
Why it matters
As a general pattern, low-cost boards that bundle sensors with on-device machine learning are lowering the barrier to hands-on AI, both in classrooms and for hobbyist accessibility projects. This is a single-build demonstration rather than a validated assistive product, but it illustrates how local TinyML inference can convert sound into immediate visual cues without sending audio to the cloud, which also helps with privacy.
Key Points
- 1WHAT: A Hackster project turns the UNIHIKER K10 board into a TinyML sound assistant that detects household noises and shows visual alerts for the hearing-impaired.
- 2WHY: It runs an Edge Impulse neural network locally on an ESP32-S3, flashing LEDs and the screen when confidence exceeds 75 percent.
- 3SO WHAT: It shows how affordable, sensor-rich education boards can put on-device AI within reach of students and accessibility tinkerers.
Scoring Rationale
A practical product-level real-time audio classifier relevant to applied audio and accessibility work; notable for practitioners but limited in scope and broader industry impact.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
