Sonic-Bench
A spatial audio reasoning benchmark with 2.8K Ambisonics clips and 28.8K QA pairs across 480 simulated scenes for localization, distance, and motion tasks.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info sonic-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
7 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Audio Flamingo 3 | 14.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Phi-4-MM-5.6B | 13.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 12.8 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| InternOmni | 12.8 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Kimi-Audio-7B-Instruct | 12.5 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen2.5-Omni-7B | 12.5 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Gemini 3.1 Pro | 9.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Sonic-Bench comprises 2,880 first-order Ambisonics clips and 28,800 target-specific question-answer pairs generated from 480 simulated indoor scenes using Habitat-Sim and Mini LibriSpeech speech data.
It evaluates dynamic source-centric spatial reasoning, including localization, distance estimation, and motion direction. The best reported model achieves 34.69% direction accuracy on the test split.