MusTBENCH - Transition-Aware Description: leaderboard
MusTBench (2026) tests temporal grounding of audio-language models in full-length music tracks, zero-shot. Transition-aware description: describe in free text the musical change around a given timestamp. Mean of METEOR against the reference and CLAP audio-text similarity, both times 100 (range -50 to 100); higher is better.
Metric: Transition-Aware Description (TAD, 208 questions): the model describes the musical change around a given timestamp in free text; score is the mean of METEOR against the reference description and CLAP score (cosine similarity of the 20-second transition clip and the generated text), both times 100: METEOR spans 0 to 100 and CLAP -100 to 100, so the mean spans -50 to 100, on the MusTBench test set (full-length MTG-Jamendo and Slakh2100 tracks, music-expert-validated questions), zero-shot with the full track and a task instruction, temperature 0, unparseable timestamps and intervals counted wrong; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 12 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 24.3 | #175 |
| 2 | Gemini 2.5 Pro | 24.1 | #276 |
| 3 | GPT-Audio 1.5 | 23.6 | #501 |
| 4 | Music Flamingo | 23.4 | #866 |
| 5 | Gemini 2.5 Flash | 22.9 | #463 |
| 6 | GPT-Audio | 22.5 | #607 |
| 7 | Qwen2.5-Omni-3B | 22.3 | #3519 |
| 8 | Qwen2.5-Omni-7B | 21.5 | #1884 |
| 9 | Gemini 3 Pro | 19.8 | #145 |
| 10 | AF-Next | 19.2 | |
| 11 | Phi-4 Multimodal Instruct | 16.3 | #4154 |
| 12 | Qwen 3 Omni 30B A3B | 16 | #847 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=mustbench-transition-aware-description · How It Works · Data refreshed daily, snapshot 2026-11-21.