HelixBench checks whether AI sound follows the camera

A fountain sounds convincing until you turn away and the water still seems to come from the wrong side. In an interactive scene, an attractive picture and a plausible soundtrack leave that question unanswered.

Noiz AI’s repository dates the public HelixBench release to October 12, 2026. It is a benchmark for generated audio and video. The associated HelixWorld preview dates back to September 3, and the paper was first posted September 29 UTC. This update concerns the benchmark, rather than a newly launched preview. Author chronology, paper history.

Direction and timing answer different questions

The public dataset card describes 1,015 fixed cases and six separate metrics. Event timing is measured on 165 event cases; spatial behavior uses 150 perspective cases. The entry point does not combine everything into one overall score.

For a creator comparing interactive sound, that means a result about left/right direction does not settle whether a footstep lands with the visible movement. Compare the same metric and cohort first, then listen for the particular failure that matters in your scene.

An illustrative check is to keep one visible sound source in view, turn the camera, then return. Listen for where the sound appears to come from and whether it changes consistently with the view. Review visible impacts separately for timing. This is a proposed listening exercise, not an evaluation we performed.

Keep the hardware claim in context

Noiz’s announcement and limitations describe HelixWorld generating video and camera-aligned stereo together. Its reported real-time speed uses an NVIDIA H800 at 768×512, with no consumer-GPU result supplied. That research result does not establish performance on your editing computer.

The benchmark is useful for asking more precise questions; it is not an independent verdict that HelixWorld wins every one of them. We have not run its public evaluation package or checked every sample. Use the published protocol for a formal comparison, and keep the small listening check separate from a benchmark score.

For ordinary generated footage rather than an explorable scene, our Runway draft-mode coverage addresses a different production workflow.