Week 32: Edge Impulse & DIY Hearing Aid
J.Gong
2025-08-02
1.54min
Week 32: Edge Impulse & DIY Hearing Aid
This week marks my second week at Tallinn Summer School! 🌞
We kicked off a project to build a hearing aid device using Arduino. To distinguish different sounds, I trained a machine learning model with Edge Impulse.
However… Arduino didn’t have enough memory 😅. So, we switched to a Heltec LoRa ESP32 — much better for this kind of project.
Project Information
| Title | Tactile feedback for deaf people |
| Authors | Gong, Yin & Nienke |
Problem
- Deaf people can’t hear alarms, phones or baby’s crying when they don’t wear their hearing aids
- They should be made aware of this using tactile feedback
Persona 1: Claire, 25 years old
| Segment | single, deaf parent of baby, living alone |
| Bio | Deaf from birth. Has just ended a long-term relationship and now has to care for her 1-year old baby alone as a first-time mom. |
| Location | Amsterdam |
| Goals | Get notified when her child needs her, without disturbing people around her |
Scenario
Persona 2: Jun Li, 84 years old
| Segment | Older people with hearing loss, living alone |
| Bio | His wife just passed away. His children live abroad. Because he can not adapt the culture abroad, he lives alone now. Since a couple of years he doesn’t hear so good anymore. He is afraid of death and his children worry about him. |
| Location | Shanghai |
| Goals | Wants to stay in his own house |
| Feel safe | |
| Get alerted when there is an emergency or phone call from his children. |
Scenario
Low Fidelity Prototype
Features:
- vibration starts only when sound exceeds threshold
- vibration stops as soon as sound stops
- device can be worn on both sides
High Fifelity Prototype
Why Frequency analysis did not work
- Phone, baby and alarm do not contain one clear frequency
- The frequencies within the same sound changed more than the frequencies between sounds
Use MFCC + CNN
The training process is open source on edge impulse.
State Machine
Some useful links from the project:
For guidance, I also watched tutorials on:
To display images on the screen, I experimented with XBM files, following this tutorial 🖼️.
It’s been a fun mix of hardware, ML, and a bit of trial and error — learning by doing is the best! 🚀