Hello! Have you ever used text to speech approaches to generate samples for training set? And what size of training set do you usually have?