AudioCaps
Models match natural language descriptions to audio clips. Meta also compares how accurately encoders match audio, video, and text from the same clip.
- Made by
- Chris Dongjoo Kim and colleagues
- Kind
- Audio understanding and voice
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision, Audio and speech
TopDown-VGGish(FC2,C4)-LSTM tops the leaderboard with 62.9%
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall score, so this is the B-1 caption similarity column, the first metric in Table 2. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.
- 62.9 TopDown-VGGish(FC2,C4)-LSTM
- 62.2 TopDown-SemTempAtt(1NN) (You et al., 2016)
- 61.8 TempAtt-VGGish(C4)-LSTM (Luong et al., 2015)
- 61.4 TopDown-AlignedAtt(1NN)
- 61.3 TempAtt-VGGish(FC2)-LSTM (Luong et al., 2015)
- 61.2 TempAtt-VGGish(C3)-LSTM (Luong et al., 2015)
- 60.9 TopDown-VGGish(FC2,C4,C3)-LSTM
- 58.7 VGGish-LSTM (Hershey et al., 2017)
- 57.7 TempAtt-MFCC-LSTM (Luong et al., 2015)
- 57.3 MFCC-LSTM (Xu et al., 2018)
- 55.5 TempAtt-SoundNet-LSTM (Luong et al., 2015)
- 54.0 SoundNet-LSTM (Aytar et al., 2016)
- 50.7 TempAtt-WaveNet-LSTM (Luong et al., 2015)
- 48.9 WaveNet-LSTM (Engel et al., 2017)
- 44.2 1NN-VGGish (Hershey et al., 2017)
- 39.1 1NN-SoundNet (Aytar et al., 2016)
- 34.1 1NN-MFCC
- 1 TopDown-VGGish(FC2,C4)-LSTM 62.9%
- 2 TopDown-SemTempAtt(1NN) (You et al., 2016) 62.2%
- 3 TempAtt-VGGish(C4)-LSTM (Luong et al., 2015) 61.8%
- 4 TopDown-AlignedAtt(1NN) 61.4%
- 5 TempAtt-VGGish(FC2)-LSTM (Luong et al., 2015) 61.3%
- 6 TempAtt-VGGish(C3)-LSTM (Luong et al., 2015) 61.2%
- 7 TopDown-VGGish(FC2,C4,C3)-LSTM 60.9%
- 8 VGGish-LSTM (Hershey et al., 2017) 58.7%
- 9 TempAtt-MFCC-LSTM (Luong et al., 2015) 57.7%
- 10 MFCC-LSTM (Xu et al., 2018) 57.3%
- 11 TempAtt-SoundNet-LSTM (Luong et al., 2015) 55.5%
- 12 SoundNet-LSTM (Aytar et al., 2016) 54%
- 13 TempAtt-WaveNet-LSTM (Luong et al., 2015) 50.7%
- 14 WaveNet-LSTM (Engel et al., 2017) 48.9%
- 15 1NN-VGGish (Hershey et al., 2017) 44.2%
- 16 1NN-SoundNet (Aytar et al., 2016) 39.1%
- 17 1NN-MFCC 34.1%
Source: The AudioCaps paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | TopDown-VGGish(FC2,C4)-LSTM | 62.9% | Official leaderboard | ||
| 2 | TopDown-SemTempAtt(1NN) (You et al., 2016) | 62.2% | Official leaderboard | ||
| 3 | TempAtt-VGGish(C4)-LSTM (Luong et al., 2015) | 61.8% | Official leaderboard | ||
| 4 | TopDown-AlignedAtt(1NN) | 61.4% | Official leaderboard | ||
| 5 | TempAtt-VGGish(FC2)-LSTM (Luong et al., 2015) | 61.3% | Official leaderboard | ||
| 6 | TempAtt-VGGish(C3)-LSTM (Luong et al., 2015) | 61.2% | Official leaderboard | ||
| 7 | TopDown-VGGish(FC2,C4,C3)-LSTM | 60.9% | Official leaderboard | ||
| 8 | VGGish-LSTM (Hershey et al., 2017) | 58.7% | Official leaderboard | ||
| 9 | TempAtt-MFCC-LSTM (Luong et al., 2015) | 57.7% | Official leaderboard | ||
| 10 | MFCC-LSTM (Xu et al., 2018) | 57.3% | Official leaderboard | ||
| 11 | TempAtt-SoundNet-LSTM (Luong et al., 2015) | 55.5% | Official leaderboard | ||
| 12 | SoundNet-LSTM (Aytar et al., 2016) | 54% | Official leaderboard | ||
| 13 | TempAtt-WaveNet-LSTM (Luong et al., 2015) | 50.7% | Official leaderboard | ||
| 14 | WaveNet-LSTM (Engel et al., 2017) | 48.9% | Official leaderboard | ||
| 15 | 1NN-VGGish (Hershey et al., 2017) | 44.2% | Official leaderboard | ||
| 16 | 1NN-SoundNet (Aytar et al., 2016) | 39.1% | Official leaderboard | ||
| 17 | 1NN-MFCC | 34.1% | Official leaderboard |
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Audio and video
- Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy.
Qwen 3.8 Max 88.7%
- LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization.
Gemini 4 Argon 91.7%
- MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy.
Qwen 3.8 Max 82.4%
- MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy.
Qwen 3.8 Max 90.8%
- FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better.
Phi-4-multimodal 4.00%
- VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules.
Qwen3.5-Omni-Plus 93.1%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the benchmark’s paper (17 entries).
- Pass Models scored 17 entries on its leaderboard, none a model we track.
- Note Newest score No scores yet.
- Note Frontier models No models scored yet.
- Note Headroom No scores yet.
Leaderboard import 3 pass
- Pass Importer We read the board on audiocaps.github.io: 17 entries at the last read, none a model we track.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.