TinyML: Wake Word Detection (jd86)

Introduction

I have the experience of do the neural network to train the images before in the AI course with CNN and RNN. Now I want to get access of the voice recognition. I learn the procedures in details from the textbook to get to set up the training model of wake words. And then I learned to modify the and edit the program so that it can recognize other words and do reaction after heard the words.

Objective

Build a small application to recognize the words "left" and "right" by training on a dataset of speech commands. Once the application listened and recognized the words, it will do reaction by lighting the LED light.

Application Architecture

• Derive the audio data set of the input for training.
• Create architecture and add model code.
• The detection responder toggles the coresponding color of the light of the model output.

Model

The project contains the model input with a dataset of Speech Commands.

• "silence"
• "unknown"
• "left"
• "right"

Swap the array of the class and match the labels with the model's output tensor elements in order.

Step

• Use Google Colaboratory service to build, train and convert the model. Load the notebook and run in Google Colaboratory.
• Derive the script of code by access to Examples -> TensorFlowLite -> micro_speech.
• Edit and save the files so that the application can recognize "left", "right", "silence" and "unknown", upload its to the board.
• Open the monitor in Arduino IDE and say word "left", "right" and any other words to see the changes of the IED light of the board.

Result