Research map / Video / Video-language understanding
HumanOmni-Speaker: Identifying Who said What and When
Detao Bai, Xihan Wei, Zhiheng Ma
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 Visual Delta Encoder (25 fps) (ablation) | Qwen2.5-Omni-SFT · Speaker Localization (easy) | miss rate (%): 3.0 -> 0.8 | Improved |
| add Visual Delta Encoder (ablation) | Qwen2.5-Omni-SFT · Speaker Identification (hard) | error rate (%): 33.2 -> 21.1 | Improved |
| add Visual Delta Encoder (ablation) | Qwen2.5-Omni-SFT · VR-SDR (hard) | SA-WER (%): 52.1 -> 47.1 | Improved |
| add Visual Delta Encoder (ablation) | Qwen2.5-Omni-SFT · VR-SDR (hard) | IER (%): 31.5 -> 28.5 | Improved |
| use Visual Delta Encoder alone without Visual Base Encoder (ablation) | HumanOmni-Speaker · Speaker Localization (easy) | miss rate (%): 3.0 -> 1.2 | Improved |
| add Visual Base Encoder to Delta Encoder (ablation) | HumanOmni-Speaker · Speaker Localization (easy) | miss rate (%): 1.2 -> 0.8 | Improved |
| end-to-end model instead of Whisper-diarization + Qwen3-VL pipeline (ablation) | HumanOmni-Speaker · VR-SDR (hard) | SA-WER (%): 64.2 -> 47.1 | Improved |
| feed full frames instead of face crops (ablation) | HumanOmni-Speaker · Speaker Identification (hard) | error rate (%): 38 -> 21 | Improved |
| add visual stream to audio | HumanOmni-Speaker · LRS3 | WER (%): 3.63 -> 0.76 | Improved |
| add visual stream to audio | HumanOmni-Speaker · LRS2 | WER (%): 3.47 -> 1.36 | Improved |
| add Delta Encoder training | HumanOmni-Speaker · LibriSpeech test-clean | WER (%): 2.2 -> 1.88 | No clear effect |
| reduce structured tokens below 6 (ablation) | Visual Delta Encoder · Speaker Localization | accuracy | Hurt |
Closest papers by meaning
- OpenOmni: Advancing Open-Source Omnimodal Large Language Models with Progressive Multimodal Alignment and Real-time Emotional Speech Synthesis (NeurIPS 2025 · similarity 0.86)
- Who Says What: Symbolic Trimodal Binding Mechanisms in Audio-Visual LLMs (NeurIPS 2026 · similarity 0.85)
- JointAVBench: A Benchmark for Joint Audio-Visual Reasoning Evaluation (ICLR 2026 · similarity 0.85)
- TalkCuts: A Large-Scale Dataset for Multi-Shot Human Speech Video Generation (NeurIPS 2025 · similarity 0.85)
- Human-MME: A Holistic Evaluation Benchmark for Human-Centric Multimodal Large Language Models (ICLR 2026 · similarity 0.85)
- Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception (ICLR 2026 · 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.
