WakeBuilder Documentation¶

What is WakeBuilder?¶

WakeBuilder is a comprehensive training platform that enables you to create custom wake word detection models entirely on your local machine. Whether you want to build a personal voice assistant that responds to "Phoenix," create a smart home automation triggered by "Hey Computer," or develop a unique voice interface for your application, WakeBuilder makes it possible—without sending your voice data to the cloud.

How It Works¶

WakeBuilder uses a powerful transfer learning approach based on the Audio Spectrogram Transformer (AST):

graph LR
    A[Your Voice<br/>1-5 Recordings] --> B[Data Augmentation<br/>+ TTS Synthesis]
    B --> C[Thousands of<br/>Training Samples]
    C --> D[AST Feature<br/>Extractor]
    D --> E[Trained<br/>Classifier]
    E --> F[Your Custom<br/>Wake Word Model]

1. Record - Speak your wake word 1-5 times into your microphone
2. Augment - WakeBuilder generates thousands of variations using multiple TTS voices
3. Train - A classifier learns to recognize your specific wake word
4. Test - Validate your model with real-time testing
5. Deploy - Export and use your model with WakeEngine

Quick Links¶

Key Features¶

🎯 Transfer Learning with AST¶

WakeBuilder uses the pre-trained MIT/ast-finetuned-speech-commands-v2 model as a frozen feature extractor. This 87-million parameter model already understands speech patterns from training on diverse audio datasets. You only train a small classifier head (~230K parameters), enabling effective learning from minimal examples.

🔊 Multi-TTS Data Augmentation¶

Generate diverse training data using multiple Text-to-Speech engines:

🛡️ Hard Negative Generation¶

WakeBuilder automatically generates phonetically similar words (like "jarv" for "jarvis") to train your model to reject near-matches, dramatically reducing false positives.

♿ Accessibility-Focused¶

WakeBuilder is especially valuable for users with unique speech patterns or pronunciation differences:

• Speech disabilities - Train a model tuned to your specific voice
• Accents - Create models that understand your natural pronunciation
• Health conditions - Works with 1-5 recordings of however you speak

🌐 Beautiful Web Interface¶

A modern, intuitive web interface guides you through the entire process:

• Model Dashboard - View, test, and manage all your wake word models
• Training Wizard - Step-by-step guidance for creating new models
• Real-time Testing - Test models with live audio from your microphone
• Training Metrics - Monitor progress with live charts and statistics

System Requirements¶

Training Time Expectations¶

Training time varies significantly based on your hardware:

Why Local Training Matters¶

• No API credits - Train unlimited models at no cost
• No cloud dependency - Works completely offline (except Edge TTS)
• Complete privacy - Your voice never leaves your machine
• Full ownership - You own your trained models completely

License¶

WakeBuilder is released under the Apache License 2.0, allowing both personal and commercial use. See the full license for details.