TinyML: Wake Word Detection

Introduction

This project was a wake word detection APP running on Sense Arduino Nano 33 BLE Sense according to Chapter 7-8 of TinYM - New Book from Google Researchers Peter Warden & Daniel Situnayake. There are two apps, the detection words are yes/no, on/off. When the board detects yes or ON, it will light up green; when it detects no or off, it will light up red; if it cannot be classified, it will light up blue.

Approach

• .Configure training

project1: set WANTED_WORDS = "yes,no", set TRAINING_STEPS = "15000,3000"

LEARNING_RATE = "0.001,0.0001"

project2: set WANTED_WORDS ="on,off", set TRAINING_STEPS = "15000,3000"

LEARNING_RATE = "0.001,0.0001"

• start training

follow Colab to Training cell and wait for one or two hour for finish training

result:

• Convert to TensorFlow Lite, then Convert to a C Array

follow Colab:

result:

• Replacing the Model

find micro_features_model.cpp, replace the DATA_ALIGN_ATTRIBUTE to the C array I have generated and change the g_model_len to the length of thenew C array.

• Updating the Labels

find micro_features_micro_model_settings.cpp, change to

const char* kCategoryLabels[kCategoryCount] = {
    "silence",
    "unknown",
    "on",
    "off",
};

• Updating command_responder.cc
• find arduino_command_responder.cpp, change to

if (found_command[0] == 'o' && found_command[1] == 'n') {
    last_command_time = current_time;
    digitalWrite(LEDG, LOW);  // Green for yes
}
if (found_command[1] == 'f') {
    last_command_time = current_time;
    digitalWrite(LEDR, LOW);  // Red for no
}

• run and test

The APP that detects yes and no has a better effect, while The effect of detecting on and off is not very good.