Skip to content

Creating Your First Model

A complete walkthrough of creating your first custom wake word model.

Before You Begin

Ensure you have:

  • WakeBuilder installed and running at http://localhost:8000
  • Negative data downloaded (check home page status)
  • Negative cache built (optional but recommended)
  • A working microphone

Step 1: Choose Your Wake Word

  1. Open WakeBuilder in your browser: http://localhost:8000
  2. Click "Train New Model" on the home page

Enter Wake Word

Enter your desired wake word in the input field:

Good Wake Word Examples

  • jarvis
  • phoenix
  • hey computer
  • listen up
  • activate

Wake Word Guidelines

Rule Description
Length 4-12 characters
Words 1-2 words maximum
Characters Letters and spaces only
Uniqueness Avoid common words like "the" or "yes"

Avoid These

  • Single letters or very short words ("hi", "go")
  • Very common words ("computer", "hello")
  • Words that sound like other common words
  • Long phrases (more than 2 words)

Step 2: Configure Training (Optional)

Click "Advanced Options" to customize training parameters:

Data Generation Settings

Setting Default Description
Target Positive Samples 5,000 Total positive training samples
Max Negative Chunks 0 (auto) From cached negative data
Negative Ratio 2.0x Real negatives per positive
Hard Negative Ratio 4.0x Similar-sounding word negatives

Training Parameters

Setting Default Description
Batch Size 32 Samples per training step
Max Epochs 100 Maximum training iterations
Early Stopping 8 epochs Patience for convergence
Learning Rate 0.0001 Step size for optimizer
Dropout 0.5 Regularization strength
Label Smoothing 0.1 Confidence calibration

Model Enhancements

Option Default Purpose
Focal Loss ✓ Enabled Better hard example handling
Self-Attention ✓ Enabled Discriminative feature focus
SE Block ✓ Enabled Channel attention
TCN Block ✓ Enabled Local pattern capture

For Your First Model

Leave all settings at their defaults. They've been tuned for optimal results.

Click "Continue to Recording" when ready.

Step 3: Record Your Voice

Recording Interface

You'll see:

  • A waveform visualizer showing audio levels
  • A record button in the center
  • A recordings list below (initially empty)
  • A recording counter showing progress

How to Record

  1. Prepare: Find a quiet environment with minimal background noise
  2. Position: Hold your device at normal speaking distance (arm's length)
  3. Record: Click and hold the record button
  4. Speak: Say your wake word clearly at normal volume
  5. Release: Let go of the button to stop recording

Recording Tips

Do This

  • Speak at your natural pace and volume
  • Keep recordings 0.5-2 seconds long
  • Record in the environment where you'll use the model
  • Vary your tone slightly between recordings

Avoid This

  • Whispering or shouting
  • Very long pauses before or after the word
  • Background noise (TV, music, conversations)
  • Speaking too fast or too slow

Review and Re-record

After each recording:

  1. Listen: Click the play button to hear your recording
  2. Evaluate: Is it clear? At normal volume?
  3. Delete: Click ✕ to remove a bad recording
  4. Re-record: Record again if needed

How Many Recordings?

Recordings Result
1 Minimum - works with TTS augmentation
3 Recommended - good variety
5 Maximum - best accuracy

Click "Start Training" when you have at least 1 recording.

Step 4: Monitor Training

Training Phases

Training progresses through several phases:

  1. Preparing data - Validation and processing
  2. Generating TTS samples - Creating voice variations
  3. Generating hard negatives - Similar-sounding words
  4. Loading negative cache - Background audio samples
  5. Training classifier - Neural network training
  6. Evaluating model - Performance assessment

During Training

You'll see:

  • Progress bar with percentage complete
  • Current phase description
  • Training metrics (loss, accuracy, F1)
  • Live charts showing training progress
  • Elapsed time counter

Training Metrics Explained

Metric What It Means
Train Loss Error on training data (lower = better)
Val Loss Error on validation data (lower = better)
Val Accuracy Correct predictions (higher = better)
Val F1 Balance of precision/recall (higher = better)

Normal Training Behavior

  • Loss decreases rapidly in early epochs
  • Loss stabilizes after 20-40 epochs
  • Early stopping may end training before max epochs

Step 5: Review Results

Performance Summary

After training, you'll see:

Metric Good Value Description
Accuracy >95% Overall correctness
F1 Score >0.90 Balanced performance
Threshold 0.5-0.7 Recommended detection sensitivity

Threshold Analysis Chart

The FAR/FRR chart shows:

  • FAR (False Accept Rate): Wrong triggers on non-wake-words
  • FRR (False Reject Rate): Missed detections of wake word

The recommended threshold balances these competing metrics.

Training Data Summary

Review what was used for training:

  • Number of recordings provided
  • Total positive samples generated
  • Total negative samples used
  • Training/validation split

Step 6: Test Your Model

Immediate Testing

  1. Click "Test Model" on the results page
  2. Grant microphone permission if prompted
  3. Start speaking your wake word
  4. Watch the detection indicator

Detection Indicator

  • Green pulse: Wake word detected
  • No response: Not detected or below threshold

Adjusting Sensitivity

Use the threshold slider to adjust detection sensitivity:

Threshold Behavior
Lower (0.3) More sensitive, may have false triggers
Higher (0.8) Less sensitive, may miss some detections

Step 7: Download and Deploy

Download Options

  • ZIP file: Contains model (.pt) and metadata (.json)
  • ONNX format: For deployment with WakeEngine

What's Included

your_wake_word_model.zip
├── your_wake_word.pt        # PyTorch model weights
├── your_wake_word.json      # Metadata and threshold
└── your_wake_word.onnx      # ONNX export (if available)

Congratulations! 🎉

You've created your first custom wake word model. Next steps: