wav2vec 2.0

wav2vec 2.0, released by Meta AI (FAIR) in September 2020, showed for the first time that self-supervised representations learned from raw speech audio alone could beat the best semi-supervised speech recognition. It masks the speech input in latent space and solves a contrastive task over jointly learned quantized representations, then fine-tunes with CTC on tiny labeled sets.

The headline result: 10 minutes of transcribed speech plus 53K hours of unlabeled audio reached competitive LibriSpeech word error rates — about 100x less labeled data than prior state of the art. The cross-lingual XLSR extension let low-resource languages piggyback on high-resource ones.

Code and pretrained models were open-sourced and it became the backbone of Meta's later MMS project and countless low-resource ASR efforts. Note: a newer Wav2Vec2-BERT 2.0 exists pretrained on 4.5M hours of audio, which HF recommends for new fine-tuning.