Skip to content

Configuration

Customize WakeBuilder's behavior through various configuration options.

Overview

WakeBuilder provides extensive configuration options for:

  • Audio processing parameters
  • Training hyperparameters
  • Data augmentation settings
  • Model architecture options
  • API configuration

In This Section

  • Audio Settings

    Sample rate, mel spectrogram parameters, and audio processing.

    Audio Settings

  • Training Hyperparameters

    Learning rate, batch size, epochs, and optimization settings.

    Hyperparameters

  • Data Augmentation

    TTS settings, speed variations, noise levels, and sample ratios.

    Augmentation

  • Model Options

    Classifier architecture, attention, SE blocks, and TCN.

    Model Options

Configuration File

All settings are centralized in src/wakebuilder/config.py:

# Audio processing parameters
AUDIO_CONFIG = {
    "sample_rate": 16000,
    "n_mels": 80,
    "n_fft": 512,
    "hop_length": 160,
    "win_length": 400,
    "fmin": 0,
    "fmax": 8000,
    "duration": 1.0,
}

# Training parameters
TRAINING_CONFIG = {
    "embedding_dim": 768,
    "hidden_dims": [256, 128],
    "dropout_rate": 0.3,
    "learning_rate": 0.001,
    "batch_size": 32,
    "max_epochs": 50,
    "early_stopping_patience": 5,
    "validation_split": 0.2,
}

Web Interface Configuration

Most settings can be configured through the Advanced Options panel in the training wizard:

Data Generation

Setting Default Description
Target Positive Samples 5,000 Total positive training samples
Max Negative Chunks Auto Cached negative data samples
Negative Ratio 2.0x Real negatives per positive
Hard Negative Ratio 4.0x Similar words per positive

Training Parameters

Setting Default Range Description
Batch Size 32 8-256 Samples per gradient update
Max Epochs 100 10-500 Maximum training iterations
Early Stopping 8 0-15 Patience epochs
Learning Rate 0.0001 0.000001-0.1 Optimizer step size
Dropout 0.5 0-0.7 Regularization strength
Label Smoothing 0.1 0-0.3 Confidence calibration
Mixup Alpha 0.5 0-1.0 Data mixing intensity

Model Enhancements

Option Default Description
Focal Loss ✓ Better hard example handling
Focal Alpha 0.5 Class weight balance
Focal Gamma 2.0 Focusing strength
Self-Attention ✓ Attention pooling layer
SE Block ✓ Channel attention
TCN Block ✓ Temporal convolutions
Classifier Layers 256, 128 Hidden dimensions

Environment Variables

Override defaults with environment variables:

# Server configuration
WAKEBUILDER_HOST=0.0.0.0
WAKEBUILDER_PORT=8000

# GPU selection
CUDA_VISIBLE_DEVICES=0

# Logging
LOG_LEVEL=info

Create a .env file in the project root to persist settings.

Quick Reference

  • Target Positives: 3,000
  • Hard Negative Ratio: 2.0x
  • Batch Size: 64
  • Max Epochs: 50
  • Early Stopping: 5
  • Target Positives: 5,000
  • Hard Negative Ratio: 4.0x
  • Batch Size: 32
  • Max Epochs: 100
  • Early Stopping: 8
  • Target Positives: 8,000
  • Hard Negative Ratio: 6.0x
  • Batch Size: 32
  • Max Epochs: 200
  • Early Stopping: 15