Finetune from a supervised pre-trained Zipformer model
======================================================