Label-efficient audio classification through multitask learning and self-supervision