# Self-supervised speech representation

A variant of Meta's Hubert model proposed in the TWIST paper, demonstrating significant value as a speech tokenizer for training SpeechLMs.

Hubert is a self-supervised learning-based speech representation model that learns joint acoustic and linguistic representations of speech through BERT-like predictive loss.

HuBERT is a self-supervised speech representation learning model that provides aligned target labels for BERT-like prediction loss through offline clustering steps, suitable for speech recognition, generation, and compression tasks.