What happens when you take the semi-supervised ASR recipe that beat the state of the art at 100M parameters and just… keep scaling? Google's BigSSL answers with an 8-billion-parameter Conformer trained on a million hours of unlabeled YouTube audio, plus tens of thousands of hours of supervised data — and shows that the gains keep coming, especially on the tail: low-resource domains, accented speech, and long-form recognition.

The paper is really two contributions in one. It's an empirical study of what breaks and what scales when you push SSL past a billion parameters, and it's a demonstration that a single pretrained encoder can act as a universal starting point for a dozen downstream ASR tasks, cutting the WER floor on YouTube, telephony, and voice search benchmarks alike. The talk covers the pretraining pipeline, the transfer results across domains, and the practical implications for anyone weighing whether a foundation-model approach makes sense for their speech stack. Worth queueing up if you're planning your next training run at scale.