6 of 6 boards
Sound
Short sounds and spoken keywords, from a microphone. Teach a board to hear a keyword, a clap, a door knock or the click of a switch, and to tell them from background noise.
Limits first
- Short sounds and a handful of keywords, not speech recognition.
- Your board needs a microphone. The Nano 33 BLE has one onboard; the other boards need a microphone module, and on some of them memory is tight (see the table).
- The downloaded code runs the model on a buffer of audio your code fills.
Which boards can do it
| Board | Support | Why |
|---|---|---|
| ESP32-S3-DevKitC-1 (N8R8) | Supported | |
| Arduino Nano 33 BLE Sense Rev2 | Supported | Onboard mic |
| Raspberry Pi Pico 2 W | Limited | Module + tight RAM |
| Nordic nRF52840 DK | Limited | Module |
| ST Nucleo-H753ZI | Supported | |
| ST Nucleo-F446RE | Limited | Module + 128 KB is tighter than nano33ble/nrf52840dk's 256 KB |
How you collect it
Record with your phone's microphone, one sound at a time or in one continuous take that is split at the silences for you. Audio is recorded at 16 kHz.
What you need
- At least 5 recordings of each sound, from at least 2 separate sessions, before the report will quote an accuracy.
- A label for the background, recorded in the room the board will live in, so the model learns what to ignore.
Getting a good model
Record each sound in several separate sessions and at different distances from the microphone. Sounds are turned into MFCC features and classified by a small depthwise-separable network that is built to fit a microcontroller.
Start a sound project
Create a free account, open a project and add your first examples.
Other signals: MotionSensor values