RAVE supports mute, compress, and gain augmentations to improve generalization on small datasets

Three training augmentations are available via --augment: mute (randomly silences data batches, default prob 0.1, teaching the model to represent silence), compress (random dynamic range compression equivalent to light non-linear amplification of batches), and gain (random amplitude scaling, default range [-6, 3] dB). These augmentations are applied on-the-fly during training. They improve generalization particularly when the dataset is small (a low-data regime), where the model might otherwise overfit to the exact loudness and noise floor of the training recordings.

Examples

rave train --config v2 --augment mute --augment compress --augment gain --db_path /data --out_path /out --name vocal_model

Assessment

A user has 45 minutes of recordings of one singer. Which augmentations are most important to add and why? What problem does the mute augmentation specifically address?