Does learning about diversity allow for good generalization?

[sangheonEN]

notebooks/training_models.ipynb

For example, if there are only male voices in the positive sample, will it work for female voices as well if data augmentation or voice synthesis is done?

[sixtyfive]

The quality of such models, generally-speaking, usually does go up with more diversity in the training data. Note that to a certain degree you can fake diversity (i.e. changing the pitch of male-only samples to produce different-sounding male or female voices as well, or overlaying various kinds of noise), but real diversity tends to work much better.