Research map / Large language models / Language, safety and interpretability
Lip Reading for Low-resource Languages by Learning and Combining General Speech Knowledge and Language-specific Knowledge
Minsu Kim, Jeong Hun Yeo, Jeongsoo Choi, Yong Man Ro
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add LMDecoder (general + language-specific knowledge) on English | AV-HuBERT · LRS2 | WER (%): 25.5 -> 23.8 | Improved |
| pre-train visual encoder on speech-unit prediction instead of supervised English lip reading | CM-seq2seq · mTEDx-FR | WER (%): 88.41 -> 77.36 | Improved |
| pre-train decoder on target-language audio-text data | AV-HuBERT · mTEDx-ES | WER (%): 71.68 -> 70.80 | Improved |
| use LMDecoder with memory instead of plain ASR-pretrained decoder | ASR Pre-train · mTEDx-IT | WER (%): 71.28 -> 68.04 | Improved |
| add Language-specific Memory (ablation) | LMDecoder · mTEDx-IT | WER (%): 71.01 -> 68.04 | Improved |
| add MLS-IT audio-text data for LMDecoder (ablation) | LMDecoder · mTEDx-IT | WER (%): 73.24 -> 70.45 | Improved |
| combine MLS-IT and mTEDx-IT audio-text data (ablation) | LMDecoder · mTEDx-IT | WER (%): 70.45 -> 68.04 | Improved |
| reduce video-text data to 15.7h (ablation) | Proposed Method · mTEDx-IT | WER (%): 68.04 -> 75.62 | Hurt |
| use only 12h audio-text data for LMDecoder (ablation) | Proposed Method · mTEDx-IT | WER (%): 73.24 -> 86.7 | Hurt |
| increase audio-text data from 147h to 294h (ablation) | Proposed Method · mTEDx-IT | WER (%): 70.3 -> 68.0 | Improved |
| LMDecoder ASR performance vs plain ASR pre-train | LMDecoder · mTEDx-IT (ASR) | WER (%): 24.65 -> 29.21 | Hurt |
| add LMDecoder to 3448h-English-pretrained encoder | CM-seq2seq · mTEDx-ES | WER (%): 59.90 -> 56.96 | Improved |
Closest papers by meaning
- Lip Reading Sentences in the Wild (CVPR 2017 · similarity 0.90)
- Hearing Lips: Improving Lip Reading by Distilling Speech Recognizers (AAAI 2020 · similarity 0.90)
- Personalized Lip Reading: Adapting to Your Unique Lip Movements with Vision and Language (AAAI 2025 · similarity 0.89)
- Language Without Borders: A Dataset and Benchmark for Code-Switching Lip Reading (NeurIPS 2024 · similarity 0.89)
- LipVoicer: Generating Speech from Silent Videos Guided by Lip Reading (ICLR 2024 · similarity 0.87)
- Let There Be Sound: Reconstructing High Quality Speech from Silent Videos (AAAI 2024 · similarity 0.85)
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.
