Research map / Vision-language and multimodal / Cross-modal learning
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, Merouane Debbah
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 |
|---|---|---|---|
| replace supervised CTC finetuning with latent-to-latent training | AV-HuBERT Large (1759h pretrain, 433h) · LRS3 test | WER: 38.6 -> 26.0 | Improved |
| replace supervised finetuning with latent-to-latent training on a Large encoder pretrained only on LRS3 433h | AV-HuBERT Large (433h pretrain, 30h) · LRS3 test | WER: 44.8 -> 55.4 | Hurt |
| use Lip2Vec instead of supervised CE finetuning on Base encoder (1759h, 433h) | AV-HuBERT Base · LRS3 test | WER: 34.8 -> 34.1 | Improved |
| add MSE loss term (alpha=0.01) to cosine loss (ablation) | Lip2Vec · LRS3 test (Large, 1759h) | WER: 34.6 -> 31.2 | Improved |
| increase MSE weight alpha beyond 0.05 (ablation) | Lip2Vec · LRS3 test (Large, 1759h) | WER: 31.2 -> 91.3 (alpha 0.5) | Hurt |
| train longer without MSE loss (ablation) | Lip2Vec alpha=0 · LRS3 test (Large, 1759h) | WER: 31.4 vs 31.2 | No clear effect |
| use progressive audio-latent masking instead of no masking (ablation) | Lip2Vec prior network · LRS3 test (Large, 1759h) | WER: 75.2 -> 31.2 | Improved |
| use progressive masking instead of fixed 100% masking (ablation) | Lip2Vec prior network · LRS3 test (Large, 1759h) | WER: 65.3 -> 31.2 | Improved |
| add 30h VoxCeleb2-en to LRS3-30h for prior training (ablation) | Lip2Vec Large · LRS3 test | WER: 31.2 -> 30.4 | Improved |
| add 30h VoxCeleb2-en to LRS3-30h for prior training (ablation) | Lip2Vec Large · VoxCeleb2-en test | WER: 39.4 -> 33.1 | Improved |
| use CTC decoding with latent mapping instead of autoregressive decoders | Lip2Vec · VoxCeleb2-en inference | runtime per 100 frames (s): 0.97 -> 0.07 | Improved |
Closest papers by meaning
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.90)
- VSRo-200: A Romanian Visual Speech Recognition Dataset for Studying Supervision and Multimodal Robustness (NeurIPS 2026 · similarity 0.89)
- Multi-Task Corrupted Prediction for Learning Robust Audio-Visual Speech Representation (ICLR 2025 · similarity 0.89)
- Restoring Speaking Lips from Occlusion for Audio-Visual Speech Recognition (AAAI 2024 · similarity 0.88)
- LipVoicer: Generating Speech from Silent Videos Guided by Lip Reading (ICLR 2024 · similarity 0.88)
- Visual Hallucination Elevates Speech Recognition (AAAI 2024 · similarity 0.88)
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.
