augment_config.yml

# Audio processing parameters
min_size_samples: 32000       # Minimum audio length (2 seconds at 16kHz)
max_size_samples: 320000      # Maximum audio length (20 seconds at 16kHz)
truncate_window_overlap_length: 16000  # Overlap when splitting long audio

# Spectrogram feature extraction
window_length_samples: 400    # Window length for STFT
hop_length_samples: 160       # Hop length for STFT
sampling_rate: 16000          # Audio sample rate
stride: 2                     # Convolution stride for feature extraction

# Label configuration
speech_label: 1               # Label for speech segments
silence_label: 0              # Label for silence segments
ignored_idx: -100             # Index to ignore in loss calculations

# Model and processing
model_id: facebook/w2v-bert-2.0  # Pre-trained model identifier
batch-size: 32                # Batch size for processing
samples-per-shard: 1024       # Samples per Parquet shard

# Augmentation parameters
seed: 1                       # Random seed for reproducibility
min-stretch-ratio: 0.8        # Minimum time stretch ratio
max-stretch-ratio: 1.5        # Maximum time stretch ratio
augment-prob: 0.4             # Probability of applying augmentation