Assignment: Build and Deploy a Keyword Spotting Model using Edge Impulse

Overview

This assignment introduces you to the exciting world of TinyML and edge AI hardware by guiding you through the process of building a keyword-spotting system using Edge Impulse. You will deploy your model on both a smartphone and an Arduino Nano 33 BLE Sense board to evaluate its performance and usability on resource-constrained hardware.

---

Objectives

1. Learn how to build and train a keyword-spotting model using Edge Impulse.
2. Deploy the trained model to both a smartphone and embedded hardware (Arduino Nano 33 BLE Sense).
3. Analyze and report on the model's performance, including trade-offs between accuracy and inference speed.

---

Steps to Complete the Assignment

Part 1: Train and Deploy Your Model

1. Create an Edge Impulse Account:

   • Go to Edge Impulse and sign up.
   • Create a new project.

2. Follow the Tutorial:

   • Complete the Edge Impulse tutorial on building a keyword-spotting model:
     Responding to Your Voice.

3. Deploy to Your Smartphone:

   • Follow the tutorial to deploy the model to your smartphone:
     Using Your Mobile Phone.

4. Bring Your Laptop:

   • Deployment is quick (less than an hour), but ensure you have your laptop with you.

---

Part 2: Deploy to Arduino Nano 33 BLE Sense

1. Required Hardware:

   • Use the Arduino Nano 33 BLE Sense board, (only 2 boards available) as part of the TinyML Arduino Learning Kit, which includes:
     • 1 Arduino Nano 33 BLE Sense board
     • 1 OV7675 Camera
     • 1 Arduino Tiny Machine Learning Shield
     • 1 USB A to Micro USB Cable
       (Use only the Arduino Nano 33 BLE Sense board for this assignment.)

2. Install the Arduino IDE:

   • Download and install the Arduino IDE: Arduino IDE.

3. Set Up Arduino IDE:

   • In Arduino IDE, navigate to Tools > Boards > Boards Manager, search for nano 33, and install Arduino Mbed OS Nano Boards.
   • Select the Nano 33 BLE Sense board from Tools > Board.

4. Prepare and Deploy the Model:

   • After training in Edge Impulse:
     • Navigate to the Deployment page.
     • Select Arduino Library and choose INT8 Quantization.
     • Click Build to download the deployment ZIP file.
   • In Arduino IDE:
     • Include the ZIP file as a library: Sketch > Include Library > Add .zip Library.
     • Exit and re-enter the Arduino IDE to ensure the library is loaded.
     • Open the example code from:
       File > Examples > YOUR_PROJECT_NAME_inferencing > nano_ble_33_sense > nano_ble_33_sense_microphone_continuous.
     • Verify and upload the code to the board.
     • Open the serial monitor (Tools > Serial Monitor) to observe outputs.
       (First-time compilation may take ~20 minutes.)

---

Part 3: Resources

• How to Deploy a Model to Arduino:
  YouTube Video
• TinyML on Arduino Using Edge Impulse:
  Cytron Blog Post (Optional: Try controlling an LED with voice commands.)

---

Deliverables

Submit a PDF file with the following:

1. Questions:
   • Does the model perform as accurately as expected on your smartphone? List a few methods to improve the model's accuracy.
   • When building a model for resource-limited hardware, how do you balance fast inference times with acceptable model accuracy? What trade-offs did you encounter?
   • Include screenshots of the training performance from step 6 of the deployment process.
2. Videos:
   • Record and provide links to:
     • The keyword-spotting model working on your smartphone.
     • The keyword-spotting model working on the embedded Arduino board.
3. Reflections:
   • Share your experience deploying the model to your smartphone and Arduino board. Mention any technical difficulties or interesting observations.

---

2. Compilation Time:
   • Be patient during the Arduino IDE's first compilation; it can take up to 20 minutes.

---

Happy experimenting!