Text this: Domain-weighted transfer learning and discriminative embeddings for low-resource speaker verification