AI Lessons
Lesson 15: Using audio classifier in our project

Purpose: Apply an exported, custom-trained audio classifier model to trigger interactive actions in a web application.

No. of Classes

1 - (Time : 1 hour 30 minutes, Laptops/desktops : 10, Students strength : 15 to 20).

Materials Required

Laptop / Desktop with Internet connection / Wi-Fi.

A working microphone and headphones or speakers.

Visual Studio Code (or code editor) and base starter project files.

Curriculum Content

3. Programming Foundations for AI

  • 3.3 Conditionals and program state
  • 3.6 User input and interaction

5. Training Custom AI Models

  • 5.6 Exporting and deploying a model

7. Designing AI-enabled Systems

  • 7.3 Integrating models into applications
Prior knowledge
  • Training custom audio models using Teachable Machine
  • Model labels, background noise, and confidence scores
  • Basic HTML/JavaScript conditional statements
Exercises

Exercise (1)


  • Export your trained audio model from Teachable Machine (or use your downloaded model files from Lesson 14).
  • Download the Base Code and open it in your code editor.
  • Link your audio model URL inside the code to replace the placeholder model Url.
  • Map each sound class label to a specific action in the web game.
  • Run the application, trigger sounds using your microphone, and verify that the application responds only when confidence thresholds are met.

Solutions


Teacher's Instruction:
  1. Briefly review Lesson 14, highlighting how classes, audio samples, and model files (model.json, metadata.json) connect to application code.
  2. Demonstrate the project integration step by step - loading the model, listening, reading the top label, and acting on it.
  3. Explain asynchronous loading (await / callbacks) so students understand why the web app needs time to load the model before listening.
  4. Demonstrate continuous audio classification (listening loops) versus discrete user interactions.
  5. Show how if / else if logic checks top prediction labels and triggers specific game movements or state changes.
  6. Emphasise setting a confidence threshold (e.g., probability > 0.85) to avoid triggering actions on background noise or partial matches.
  7. Discuss real-world applications of an audio classification system.
  8. Points to Ponder:
    • Smart speakers wait for a "wake word" (like Hey Siri, Alexa or OK Google) before processing commands. How does a device listen for a wake word locally on your hardware without constantly sending all your private home conversations to cloud servers? Where should we draw the line between convenience and privacy?