Research map / Vision-language and multimodal / Cross-modal learning
How To Design a Three-Stage Architecture for Audio-Visual Active Speaker Detection in the Wild
Okan Köpüklü, Maja Taseska, Gerhard Rigoll
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 2D-ResNet-18 video backbone with 3D-ResNet-18 (ablation) | audio-visual encoder · AVA-ActiveSpeaker val (encoder only, 8 frames) | mAP (%): 79.0 -> 83.9 | Improved |
| replace MFCC 2D-ResNet-18 audio backbone with SincDSNet (ablation) | audio-visual encoder · AVA-ActiveSpeaker val (encoder only, 8 frames) | mAP (%): 79.0 -> 80.8 | Improved |
| use both 3D-ResNet-18 and SincDSNet (ablation) | audio-visual encoder · AVA-ActiveSpeaker val (encoder only, 8 frames) | mAP (%): 79.0 -> 86.1 | Improved |
| use 3D-ResNeXt-101 instead of 3D-ResNet-18 (ablation) | video backbone · 8-frame clips | mAP (%): 86.1 -> 86.7 | Improved |
| increase clip length from 8 to 16 frames (ablation) | 3D-ResNeXt-101 encoder · AVA-ActiveSpeaker val | mAP (%): 86.7 -> 88.9 | Improved |
| increase clip length from 16 to 32 frames (ablation) | 3D-ResNeXt-101 encoder · AVA-ActiveSpeaker val | mAP (%): 88.9 -> 88.9 | No clear effect |
| add temporal modeling to encoder (ablation) | ASDNet · AVA-ActiveSpeaker val (component analysis) | mAP (%): 88.9 -> 92.6 | Improved |
| add ISRM to encoder (ablation) | ASDNet · AVA-ActiveSpeaker val (component analysis) | mAP (%): 88.9 -> 89.6 | Improved |
| add ISRM on top of encoder plus temporal modeling (ablation) | ASDNet · AVA-ActiveSpeaker val (component analysis) | mAP (%): 92.6 -> 93.5 | Improved |
| use NonLocal instead of ISRM with temporal model (ablation) | ASDNet · AVA-ActiveSpeaker val (ISRM comparison) | mAP (%): 92.8 -> 93.4 | Improved |
| use background speaker features from 9 neighbouring frames (ablation) | ISRM · AVA-ActiveSpeaker val | mAP (%): 93.4 -> 93.5 | Improved |
| use bidirectional GRU instead of unidirectional GRU (ablation) | temporal model · AVA-ActiveSpeaker val (temporal model comparison) | mAP (%): 92.8 -> 93.5 | Improved |
Closest papers by meaning
- LoCoNet: Long-Short Context Network for Active Speaker Detection (CVPR 2024 · similarity 0.91)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.83)
- $\texttt{AVROBUSTBENCH}$: Benchmarking the Robustness of Audio-Visual Recognition Models at Test-Time (NeurIPS 2025 · similarity 0.83)
- Who Says What: Symbolic Trimodal Binding Mechanisms in Audio-Visual LLMs (NeurIPS 2026 · similarity 0.83)
- Seeing Speech and Sound: Distinguishing and Locating Audio Sources in Visual Scenes (CVPR 2025 · similarity 0.83)
- Enhancing Audiovisual Speech Recognition Through Bifocal Preference Optimization (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.
