Wake word detect

Overview

This project is followed the intructions in CHAPTER 8 of TinyML book. After training on Google Colab flatform, new model is deployed on Arduino Nano 33 BLE board. When detecting different words, different LED light would be lit accordingly.

Train

On the Google Colab platform, first of all, hardware accelerator is set to GPU in Runtime setting.

Then I simply adopt the default training parameters of Colab before training the model.

Training these words: on,off 
Training steps in each stage: 12000,3000 
Learning rate in each stage: 0.001,0.0001 Total number of training steps: 15000

To save time on training, I use the pretrained model provided by Google. The accuracy of the model is a little above 50%.

Deploy

It only take three simply steps to deploy the trained model to Arduino board.

• Generate a TensorFlow Lite for MicroControllers Model
• Replace the model and labels in original library with new model.
• Click the 'upload' button in Arduino software.

Result

Considering I keep repeating saying the same word, the behavior of microprocessor is a little poor.

Conclusion

The microprocessor on the Arduino Nano 33 BLE board is able to detect words "on" and "off" based on the trained model, though it is not sensitive to target words.