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 supervised phase can quietly wash out useful self-supervised structure. Google's JUST framework fuses them into a single joint objective — contrastive loss, masked prediction loss, and RNN-T supervised loss all backpropagating through one shared encoder at the same time.

On the 42-language Multilingual Librispeech benchmark, JUST outperforms both cascaded pretraining pipelines and pure supervised baselines, with especially strong gains on low-resource languages where labeled data is the bottleneck. The talk explains how the three losses are balanced, why joint training generalizes better across scripts and phonologies, and where the approach still leaves room for language-specific adapters. If you're wrangling a multilingual ASR stack and burning cycles on staged training runs, this one-pot recipe is worth a serious look.