Noiz AI and the Hong Kong University of Science and Technology released HelixBench on Oct. 12 to test whether generated stereo audio places sound on the same side as a visible source when a virtual camera moves. Its spatial measure compares a source's left-right position on screen with the balance between the left and right audio channels; that measure covers 150 samples, not the full 1,015-sample benchmark.
An interactive world model turns an image into a scene a user can steer. The related HelixWorld model generates video and stereo audio together as keyboard inputs move the camera. That lets the generator condition sound on the changing viewpoint instead of adding a soundtrack after the video has already been made.
How the spatial test works
HelixBench contains 1,015 human-verified samples across six scene domains, but not every measure uses all of them. The benchmark reports six separate measures rather than one combined score. Its spatial-audio measure applies to 150 samples in the Perspective subset.
For that measure, the evaluator compares where a visible sound source appears horizontally in the generated video with the left-to-right balance of the generated stereo channels. It tests whether a sound points toward the side where its source appears. That is a narrower question than whether a model reproduces a room's full acoustics or every physical property of a sound.
HelixBench's release and paper context
The frozen dataset package supplies a first frame and an audio-visual caption for each sample ID, along with the evaluation code. It does not provide complete source clips or model-generated outputs. A team evaluating its model must generate its own video and stereo audio for the test IDs in the required format, then run the released metrics.
The benchmark paper appeared on arXiv on Sept. 29, and the project repository dates the release of preview inference code and a checkpoint to Sept. 3. The Oct. 12 update adds the public HelixBench package to an earlier HelixWorld model preview and paper. The paper's benchmark comparisons are the researchers' own; the dataset release lets outside teams test the system but does not independently reproduce those results.
Sources
- Noiz AI’s HelixWorld announcement, Oct. 12
- HelixWorld paper, arXiv v1, Sept. 29
- HelixBench dataset card and evaluation protocol
- HelixWorld project repository and release history
Image: HelixWorld project banner, NoizAI/HelixWorld, Apache License 2.0. Illustrative model image, not a benchmark result.