Wat priin
Type: Course project: 'ITTT'
Learning goals: Arduino, machine learning, voice recognition
Software used: Arduino, Edge Impulse, Blender
Key takeaways: A big portion of the available time in a project goes to testing, adjusting and tweaking in a physical/real-life interactive product. Getting the technical part to work is only the first step in creating a product that actually functions in the real world, since there are so many human and worldy factors that a digital system doesn't automatically take into account.
Wat Priin was a project I did during my first year of Game Design. Its goal was to create an interactive object/artifact/installation using Arduino. For this project, I chose to dive deeper into the world of speech recognition and machine learning. I wanted to create an arduino that was able to recognize custom commands, without relying on existing databases or premade keywords. I chose to train a model using Edge Impulse to recognize two commands in a fictional language, 'turn on' and 'turn of', for which I chose the phrases 'Sɛɛwen' and 'Ad'angwen'. Using Google's existing speech recognition techniques and Python, I generated a few thousand variations of the database of samples I recorded.
Using Edge Impulse and machine learning, I taught a model to recognize the two keywords with 96.91% accuracy. I uploaded the model to an Arduino Sense and modeled and printed a casing around it.
The toughest challenge during this project was calibrating the model to work well with the Arduino and its software. The model demanded a lot of power and resources, and the Arduino Sense didn't have much available. I had to recreate and resize the model a few times to get it working. Apart from that, getting the sensitivity and error-margin right was a tedious process. The model worked accurately, but the built-in microphone of the Arduino had to be adjusted just right to be able to pick up and match the voice with the kewords. Working with real-life objects requires extensive testing and tweaking, especially when comparing it to digital products.
