Text this: Discriminative Universal Background Model Training for Speaker Recognition