hubert

Definition

Hidden-Unit BERT, a self-supervised speech representation model from Meta. Speech has no ready-made discrete units to mask and predict, so HuBERT manufactures them: cluster acoustic features offline, treat the cluster IDs as pseudo-labels for a masked-prediction objective, then re-cluster using the improved representations and repeat. The learned units proved useful well beyond low-resource ASR — they are a common tokeniser for speech generation and voice conversion.