Research map / Label-efficient and robust learning / Self-supervised and contrastive
vq-wav2vec: Self-Supervised Learning of Discrete 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 |
|---|---|---|---|
| add BERT base pre-training to Gumbel-quantized speech representations | vq-wav2vec Gumbel · WSJ nov92, No LM | WER: 14.67 -> 9.39 | Helped |
| add BERT base pre-training to Gumbel-quantized speech representations | vq-wav2vec Gumbel · WSJ nov92, 4-gram LM | WER: 6.10 -> 3.62 | Helped |
| add BERT small pre-training to Gumbel-quantized speech representations | vq-wav2vec Gumbel · TIMIT test | PER: 17.78 -> 11.64 | Helped |
| use span masking rather than masking individual tokens (ablation) | BERT trained on discretized speech · TIMIT | PER: 17.38 -> 15.28 | Helped |
| increase span length from M=10 to M=20 (ablation) | BERT trained on discretized speech · TIMIT dev | PER: 12.65 -> 13.04 | Hurt |
| increase span length from M=10 to M=30 (ablation) | BERT trained on discretized speech · TIMIT dev | PER: 12.65 -> 13.18 | Hurt |
| increase mask probability from p=0.015 to p=0.050 at fixed M=10 (ablation) | BERT trained on discretized speech · TIMIT dev | PER: 12.65 -> 11.45 | Helped |
| use multiple quantization groups instead of one group (ablation) | vq-wav2vec k-means · TIMIT dev, Libri100, V=320 | PER: 27.22 -> 20.86 | Helped |
| use a large codeword inventory (39.9M codewords) (ablation) | vq-wav2vec k-means · WSJ nov92, No LM | WER: 15.17 -> 12.17 | Helped |
| replace Gumbel quantization with k-means quantization (ablation) | vq-wav2vec without BERT · WSJ nov92, No LM | WER: 14.67 -> 15.17 | Hurt |
Closest papers by meaning
- wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (NeurIPS 2020 · similarity 0.90)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.84)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.83)
- MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training (ICLR 2024 · similarity 0.82)
- Multichannel AV-wav2vec2: A Framework for Learning Multichannel Multi-Modal Speech Representation (AAAI 2024 · similarity 0.82)
- SelfVC: Voice Conversion With Iterative Refinement using Self Transformations (ICML 2024 · similarity 0.82)
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.