Camera Rotation
Infer how the camera rotates from synchronized audio and visual cues.
OMNIECHOBENCH-QA
How well can a model understand the world through sound and sight?
Compare Audio & Vision performance across five spatial reasoning tasks.
One unified Audio & Vision table
Qwen3.8-omni-flash
FOA spatial audio + vision
OmniEcho
| # | |||||||
|---|---|---|---|---|---|---|---|
| 01 | Qwen3.8-omni-flash | 74.1 | 47.9 | 28.6 | 37.4 | 39.5 | 43.9 |
| 02 | Gemini3.8-flash | 53.7 | 34.1 | 22.1 | 31.3 | 26.0 | 32.0 |
| 03 | Gemini3.7-flash | 45.3 | 32.8 | 24.4 | 33.4 | 26.2 | 31.4 |
| 04 | muse-spark-1.2 | 55.0 | 34.2 | 19.6 | 29.6 | 25.2 | 31.3 |
| 05 | OmniEcho OURS Base: Qwen3-Omni-30B-A3B | 41.7 | 22.6 | 14.2 | 24.7 | 47.5 | 28.5 |
| 06 | doubao-seed-2-0-lite-260428 | 34.3 | 30.6 | 22.1 | 29.6 | 25.3 | 28.3 |
| 07 | Qwen3.5-Omni-Plus | 52.4 | 16.1 | 19.6 | 22.2 | 25.2 | 23.3 |
| 08 | Qwen3-Omni-30B-A3B Base: Qwen3-Omni-30B-A3B | 32.7 | 16.3 | 7.7 | 20.9 | 22.3 | 18.5 |
| 09 | Qwen2.5-Omni-7B Base: Qwen2.5-Omni-7B | 24.9 | 8.0 | 5.4 | 14.0 | 20.0 | 12.6 |
| 10 | SO-7B Base: Qwen2.5-Omni-7B | 21.4 | 5.2 | 0.0 | 17.1 | 22.8 | 11.4 |
Reading the results. All entries use audio and vision. Overall is reproduced as reported, rather than calculated as a simple mean of the five columns. Ranks remain based on Overall when sorting other metrics.
Data provenance. The Audio & Vision results from the OmniEcho paper are combined with additional results supplied by the project maintainer. OmniEcho uses the paper values (41.7 Camera Rotation; 14.2 3D Localization). Download the full data and notes ↗
BEHIND THE NUMBERS
Infer how the camera rotates from synchronized audio and visual cues.
Identify where a sound comes from, including sources outside the frame.
Estimate a sounding source’s position in three-dimensional space.
Reason about the heading and motion of a source over time.
Connect heard directions to positions on a bird’s-eye map.
Real-world spatial audio, a dedicated FOA encoder, and embodied reasoning.