Wio Terminal TinyML - Speech Recognition

The Wio Terminal is a development platform by Seeed Studio (https://www.seeedstudio.com/Wio-Termi.... It combines a SAMD51 microcontroller with Bluetooth/WiFi connectivity, various sensors, an LED screen, a buzzer, and some buttons. While it makes for a fine IoT development kit, I wanted to see how well it would run embedded machine learning applications (TinyML). Specifically, I wanted to test its ability to handle speech recognition (or more specifically, keyword spotting).

Disclosure: Seeed Studio sent me a Wio Terminal to play with and create a video about.

I trained a convolutional neural network to identify 4 keywords: forward, left, right, stop. These words were chosen to demonstrate what you might find on something like a robot. However, any words could have been chosen (within some limits). It seems that the neural network gets a little too big or inaccurate after about 5 different keywords.

You can train your own keyword spotting system by following the steps I outlined in this GitHub repository: https://github.com/ShawnHymel/ei-keyw...

The code for this demo can be found here: https://github.com/ShawnHymel/ei-keyw...

Note that you will need to install the Edge Impulse Arduino library (.zip file) found here: https://github.com/ShawnHymel/ei-keyw...

If you train your own model on Edge Impulse (https://www.edgeimpulse.com/), you can download and install the Arduino library containing your custom model. That will allow you to identify keywords of your choosing.

I recommend checking out my Coursera class on embedded machine learning if you'd like to learn more: https://www.coursera.org/learn/introd...