Research map / Label-efficient and robust learning / Self-supervised and contrastive
wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train on unlabeled speech and fine-tune with limited transcripts | wav2vec 2.0 Large on LibriSpeech · LibriSpeech test sets, 10 min labeled | WER: 16.3 -> 4.8 (test-clean); 25.2 -> 8.2 (test-other) | Helped |
| Use continuous latent inputs with quantized contrastive targets (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 12.18 -> 7.97 | Helped |
| Replace quantized targets with continuous targets (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.58 | Hurt |
| Add an MLP to continuous targets (ablation) | Continuous-input, continuous-target variant · LibriSpeech dev-clean + dev-other, train-10h | WER: 8.58 -> 8.51 | No clear effect |
| Use separate encoders for input and target representations (ablation) | Continuous-input, continuous-target variant · LibriSpeech dev-clean + dev-other, train-10h | WER: 8.58 -> 8.90 | Hurt |
| Remove the codebook diversity penalty (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.48 | Hurt |
| Double the relative positional convolution kernel (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.14 | Hurt |
| Stop quantizer gradients from reaching the encoder (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.41 | Hurt |
| Increase same-utterance distractors from K=100 to K=200 (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.12 | No clear effect |
| Sample distractors across utterances in the batch (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.79 | Hurt |
| Remove Gumbel noise from quantization (ablation) | wav2vec 2.0 · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.73 | Hurt |
| Increase the mask span length from the baseline to M=15 (ablation) | wav2vec 2.0 pre-training · LibriSpeech dev-clean + dev-other, train-10h | WER: 7.97 -> 8.43 | Hurt |
Closest papers by meaning
- data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language (ICML 2022 · similarity 0.92)
- Unsupervised Speech Recognition (NeurIPS 2021 · similarity 0.90)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.89)
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.87)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.87)
- ES3: Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.