Model Training

Training Workflow

Dataset
    |
    v
Load Audio Files
    |
    v
Extract Features
    |
    v
Create Feature Matrix X
    |
    v
Create Labels y
    |
    v
Train/Test Split
    |
    v
Train Candidate Models
    |
    v
Evaluate
    |
    v
Save Selected Model

Feature Matrix

The feature matrix contains one row per audio sample.

X = [n_samples, n_features]

For the proposed expanded feature configuration:

X = [n_samples, 191]

The actual feature count must match the finalized extraction implementation.

Target Labels

y = [REAL, FAKE, REAL, ...]

Labels must be assigned consistently and verified before training.

Dataset Balance

Class imbalance can cause a classifier to favor the majority class.

For example, if most training samples belong to REAL, a model may learn to predict REAL frequently without learning reliable distinguishing patterns.

Recommended Checks

Before training, calculate:

• Number of REAL samples
• Number of FAKE samples
• Class proportions
• Per-class validation performance

Important

Adding more samples does not automatically improve model quality.

The source, authenticity, diversity, and recording conditions of the samples also matter.