VocalGrad
VocalGrad: Evaluating Acoustic Perception in Audio Language Models
A 21,600-clip benchmark evaluating LALMs' acoustic perception via binary classification of temporal change direction in single acoustic factors.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info vocalgrad
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
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Kimi-Audio | 51.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| MiMo-Audio | 50.3 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Gemini 3 Flash Preview | 50.2 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Step-Audio 2 mini | 50.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Audio Flamingo 3 | 49.6 | N/A | Independent | medium | Oct 16, 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
VocalGrad isolates low-level acoustic perception by applying temporal transformations to single features (e.g., pitch, noise) in speech clips. The task is a simple binary classification of whether the feature increases or decreases, minimizing semantic reasoning.
Scoring uses accuracy on this binary task. Results show humans achieve near-perfect accuracy, while state-of-the-art LALMs perform at near-chance levels, revealing significant gaps in acoustic modeling capabilities.