wav2vec 2.0 changed the game with contrastive self-supervision, but HuBERT asked a sharper question: what if we skip the contrastive trick entirely and just do BERT-style masked prediction on speech? The catch is that audio has no natural vocabulary, so HuBERT bootstraps its own — clustering MFCCs first, then iteratively re-clustering the model's own hidden states — and trains the transformer to predict those pseudo-labels on masked frames.

The payoff is a cleaner objective and, on Librispeech and Libri-light, results that match or beat wav2vec 2.0 at every model scale from Base to X-Large. This deep dive from Meta AI walks through the k-means bootstrapping loop, why acoustic units emerge from iteration, and how the same pretraining transfers cleanly to ASR, speaker ID, and emotion tasks. HuBERT has since become one of the default starting points on HuggingFace for anyone doing speech representation work, so if you're picking an encoder for your next ASR or SLU project, spending a few minutes on the intuition here will save you hours of hyperparameter guessing later.