Model Architecture Options¶

Configure the classifier architecture and enhancements.

Overview¶

WakeBuilder's model consists of:

1. Frozen AST base - Extracts audio embeddings (not configurable)
2. Trainable classifier - Detects wake word (configurable)

Classifier Architecture¶

Hidden Dimensions¶

Default: [256, 128]

Sizes of fully-connected layers:

More layers = more parameters = slower training.

Dropout Rate¶

Default: 0.5

Applied between layers:

Higher = more regularization, prevents overfitting.

Enhancement Modules¶

Self-Attention Pooling¶

Default: Enabled

Learns to weight important time frames:

Parameters: ~2.4M

Recommendation: Keep enabled for best accuracy.

Squeeze-and-Excitation (SE) Block¶

Default: Enabled

Channel attention mechanism:

Parameters: ~150K

Recommendation: Keep enabled.

Temporal Convolutional Network (TCN)¶

Default: Enabled

Captures temporal patterns:

Parameters: ~1.2M

Recommendation: Keep enabled for complex wake words.

Parameter Count¶

With All Enhancements (Default)¶

Minimal Configuration¶

Without enhancements:

Configuration in Web UI¶

The Advanced Options panel includes toggles for:

• ✅ Self-Attention Pooling
• ✅ SE Block
• ✅ TCN Block

And fields for:

• Classifier Layers: 256, 128

Configuration via API¶

{
  "use_attention": true,
  "use_se_block": true,
  "use_tcn": true,
  "classifier_dims": "256,128"
}

Presets¶

Minimal (Fast)¶

{
    "classifier_dims": [128],
    "use_attention": false,
    "use_se_block": false,
    "use_tcn": false,
}

Use when: Speed is critical, simple wake words.

Balanced (Default)¶

{
    "classifier_dims": [256, 128],
    "use_attention": true,
    "use_se_block": true,
    "use_tcn": true,
}

Use when: Best accuracy/speed trade-off.

Maximum Capacity¶

{
    "classifier_dims": [512, 256, 128],
    "use_attention": true,
    "use_se_block": true,
    "use_tcn": true,
}

Use when: Complex wake words, ample training data.

Trade-offs¶

More Capacity¶

Pros:

• Can learn complex patterns
• Higher potential accuracy
• Better discrimination

Cons:

• Slower training
• Larger model size
• Risk of overfitting

Less Capacity¶

Pros:

• Faster training
• Smaller model
• Less likely to overfit

Cons:

• May miss subtle patterns
• Lower maximum accuracy
• Less robust

Recommendations¶

For Most Users¶

Use the default configuration. It balances accuracy and performance.

For Simple Wake Words¶

Disable TCN if wake word is short (1-2 syllables).

For Complex Wake Words¶

Keep all enhancements enabled. Consider larger classifier dims.

For Resource-Constrained Deployment¶

Use minimal configuration and quantized ONNX export.