Preprint
wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
Alexei Baevski et al., arXiv (NeurIPS 2020), 2020
Read the originalRead on October 16, 2026
- What it is
- The paper that introduced wav2vec 2.0, a model that learns speech representations from unlabeled audio and is then fine-tuned on a small amount of transcribed speech. It reports word error rates on LibriSpeech with full and tiny amounts of labeled data.
- How much weight it carries
- Written by the Facebook AI Research team that built the model and widely cited and reproduced. Results are on LibriSpeech read English, so they say little about conversation or noisy audio.
Facts taken from this source
- With all of LibriSpeech's labeled training data, wav2vec 2.0 reached word error rates of 1.8 percent on the clean test set and 3.3 percent on the harder other test set.
“Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets.” (Abstract)
- With just ten minutes of labeled data and pre-training on 53,000 hours of unlabeled audio, wav2vec 2.0 still reached 4.8 percent word error rate on LibriSpeech clean and 8.2 percent on other.
“Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER.” (Abstract)