Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-supervised learning with random-projection quantizer for speech recognition
Chung-Cheng Chiu, James Qin, Yu Zhang, Jiahui Yu, Yonghui Wu
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use BEST-RQ instead of wav2vec 2.0 for streaming pre-training | Conformer 0.6B · LibriSpeech test | WER: 2.9 -> 2.8 | Improved |
| Use BEST-RQ instead of w2v-BERT for streaming pre-training | Conformer 0.6B · LibriSpeech test-other | WER: 8.1 -> 6.6 | Improved |
| Use BEST-RQ streaming pre-training rather than the Conformer 0.1B baseline | Conformer 0.6B · LibriSpeech | relative latency (ms): 0 -> -130.9 | Improved |
| Use BEST-RQ instead of w2v-BERT | Conformer 0.6B · MLS-10hrs | average WER: 9.9 -> 9.6 | Improved |
| Use BEST-RQ instead of w2v-BERT | Conformer 0.6B · Voice Search VS-1000hrs | average WER: 11.5 -> 10.9 | Improved |
| Replace random quantization with Projection VQ-VAE targets (ablation) | BEST-RQ pretrain-finetune · LibriSpeech | WER dev-other: 2.8 -> 2.8 | No clear effect |
| Replace random quantization with Transformer VQ-VAE targets (ablation) | BEST-RQ pretrain-finetune · LibriSpeech | WER dev-other: 2.8 -> 2.9 | No clear effect |
| Use Transformer VQ-VAE quantized tokens rather than random quantized tokens as direct ASR inputs | Direct ASR · LibriSpeech 960h | WER dev-other: 78.8 -> 35.8 | Improved |
| Increase pre-training data when using the random-projection quantizer (ablation) | BEST-RQ · LibriSpeech fine-tuning | WER | Improved |
Closest papers by meaning
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.87)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.86)
- Multi-resolution HuBERT: Multi-resolution Speech Self-Supervised Learning with Masked Unit Prediction (ICLR 2024 · similarity 0.85)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.84)
- Speech Self-Supervised Learning Using Diffusion Model Synthetic Data (ICML 2024 · similarity 0.84)
- Heuristic-free Knowledge Distillation for Streaming ASR via Multi-modal Training (AAAI 2025 · similarity 0.83)
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.
