Tutorials
Joint Unsupervised and Supervised Training for Multilingual ASR
Pretrain-then-finetune has become the default recipe for multilingual ASR, but it leaves a lot on the table: the two stages don't share losses, and the supervi…