Research map / Vision-language and multimodal / Cross-modal learning
Efficient Audio-Visual Speech Separation with Discrete Lip Semantics and Multi-Scale Global-Local Attention
Kai Li, Gao Kejun, Xiaolin Hu
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 global attention (GA) to the GLA block (ablation) | Dolphin separator · LRS2 | SI-SNRi (dB): 15.6 -> 16.8 | Improved |
| add local attention (LA) to the GLA block (ablation) | Dolphin separator · LRS2 | SI-SNRi (dB): 15.9 -> 16.8 | Improved |
| add both GA and LA to a plain depthwise-conv block (ablation) | Dolphin separator · LRS2 | SI-SNRi (dB): 10.4 -> 16.8 | Improved |
| use HDA layer instead of 1D large-kernel convolution (ablation) | Dolphin · LRS2 | SI-SNRi (dB): 16.5 -> 16.9 | Improved |
| use discrete (VQ) lip features instead of continuous autoencoder | video encoder · LRS2 | SI-SNRi (dB): 15.2 -> 16.8 | Improved |
| add audio-aligned semantic distillation to LipCoder (DP-LipCoder) | video encoder · LRS2 | SI-SNRi (dB): 16.3 -> 16.8 | Improved |
| replace 3D ResNet-18 with DP-LipCoder | video encoder · LRS2 | SI-SNRi (dB): 17.0 -> 16.8 | No clear effect |
| replace original video encoder with DP-LipCoder in IIANet | IIANet · LRS2 | SI-SNRi (dB): 16.0 -> 15.4 | Hurt |
| replace original video encoder with DP-LipCoder in CTCNet | CTCNet · LRS2 | SI-SNRi (dB): 14.3 -> 14.0 | Hurt |
| use early audio-visual fusion in the separator encoder instead of later fusion (ablation) | Dolphin · LRS2 | SI-SNRi (dB): +0.6 dB | Improved |
| add VQ module to the video encoder (ablation) | DP-LipCoder · LRS2 | SI-SNRi (dB): approximately 0.5 dB | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Visualvoice — VisualVoice: Audio-Visual Speech Separation With Cross-Modal Consistency (CVPR 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation (ICLR 2024 · similarity 0.89)
- Restoring Speaking Lips from Occlusion for Audio-Visual Speech Recognition (AAAI 2024 · similarity 0.86)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.86)
- Seeing Speech and Sound: Distinguishing and Locating Audio Sources in Visual Scenes (CVPR 2025 · similarity 0.86)
- Learning Visual Speech Representations via Cross-Modal Distillation and Joint Face-Lip Modeling (NeurIPS 2026 · similarity 0.86)
- Improving Audio-Visual Segmentation with Bidirectional Generation (AAAI 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.
