Wake word detection Tinyml (fw18)

Introduction

“Hey Siri” was one of the most annoying but fun things on iphones when it was first developed. It is annoying when you trigger it during class and everyone starts looking at you. But the voice recognition that Siri can accomplish makes it fun and useful. That’s why I wish to accomplish a simple voice recognition - wake word detection for this project. However to do such a thing on a board is very challenging since the resource is limited. Thanks to tensorflow lite for making it possible by making the models lighter without losing much accuracy.

I will be trying to make the board recognize “yes” and “no” as wake words.

Methods

First we need to find a suitable dataset. Thankfully there are plenty on the internet today since voice recognition is a well researched area today. I used google speech command dataset “http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz”. After I found a good dataset, I used google Collab and tensorflow lite for training which was just a matter of time and changing params.

Something new in TinyMl is to convert to binary and deploy to board. This is something that I have never done before. With the Arduino IDE they worked pretty well together.

Conclusion

We were about to recognize “yes” and “no” using the board. The amount of knowledge I learned about hardware and limited resource machine learning is unbelievable. Microprocessors are extremely fun to play with. In the future I would like to try the shake to wake feature. Like when we pick up our phone it turns on by itself. I think it’s a nice feature.