Text this: Multi-scale and cardinality-based backbone: toward learning a general-purpose speaker representation