Data Augmentation Settings¶

Configure TTS and audio augmentation for training data generation.

Overview¶

Augmentation settings control how training data is generated and varied.

TTS Settings¶

Enabling TTS Providers¶

Voice Selection¶

Voices are selected automatically based on:

• Language match
• Quality tier (prefer medium/high)
• Diversity (different genders, accents)

Sample Generation¶

Target Counts¶

Sample Distribution¶

With defaults (5,000 positives):

Speed Augmentation¶

Speed Distribution¶

Pitch Augmentation¶

Pitch Values¶

Volume Augmentation¶

Volume Levels¶

Noise Augmentation¶

Noise Types¶

Configuration Example¶

AUGMENTATION_CONFIG = {
    # TTS
    "use_tts_positives": True,
    "use_hard_negatives": True,

    # Sample counts
    "target_positive_samples": 5000,
    "negative_ratio": 2.0,
    "hard_negative_ratio": 4.0,

    # Speed
    "speed_normal_range": [0.9, 1.1],
    "speed_fast_range": [1.3, 1.5],

    # Pitch
    "pitch_range": [-2, 2],

    # Volume
    "volume_range": [0.7, 1.3],

    # Noise
    "noise_snr_range": [-20, -5],
    "noise_probability": 0.5,
}

Trade-offs¶

More Augmentation¶

Pros:

• Better generalization
• More robust model
• Handles variations better

Cons:

• Longer training time
• More memory usage
• Diminishing returns

Less Augmentation¶

Pros:

• Faster training
• Lower memory usage
• Simpler data

Cons:

• May overfit
• Less robust
• Narrower detection range

Recommendations¶

Standard Use¶

Use defaults. They balance quality and training time.

Limited Hardware¶

{
    "target_positive_samples": 3000,
    "negative_ratio": 1.5,
    "hard_negative_ratio": 2.0,
}

Maximum Quality¶

{
    "target_positive_samples": 8000,
    "negative_ratio": 3.0,
    "hard_negative_ratio": 6.0,
}