---
extra_gated_prompt: >-
  Please login HuggingFace to register your email and research affiliation to
  get auto-approval. Welcome to DCASE 2025 Task-5
  https://dcase.community/challenge2025/
extra_gated_fields:
  Full Name: text
  Email: text
  Affiliation: text
  Country: country
  Specific date: date_picker
  I want to use this model for:
    type: select
    options:
    - Research
    - Education
    - Product Development
    - Others
    - label: Other
      value: other
language:
- en
license: mit
---

## Audio SFT / Post-Training Data 

<p align="center">
  <img src="img-dcase.png" width="45%">
  <br />
  <span>
  The proposed audio question answering (AQA) dataset
with three categories: Bioacoustics QA (BQA), Temporal Soundscapes QA (TSQA), and Complex QA (CQA)
  </span>
</p>

- [DCASE 2025 Task Description](https://dcase.community/challenge2025/task-audio-question-answering)
- [Audio QA Model Baseline](https://huggingface.co/PeacefulData/2025_DCASE_AudioQA_Baseline)
- [Watkins Marine Mammal Sound Database](https://whoicf2.whoi.edu/science/B/whalesounds/index.cfm)
  - "Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution and the New Bedford Whaling Museum."

---

### 📢 Post-Challenge Research Note
While the DCASE 2025 Challenge has concluded its official submission phase, this repository remains open for ongoing research. **Researchers are encouraged to continue evaluating their models and reporting results on the Development Set.** For benchmarking purposes, please refer to the baseline results provided below.

### Official Development Set Baseline Results
The following table represents the baseline performance on the Development Set as provided by the DCASE 2025 Task 5 organizers.

| Metric | BQA (Bioacoustics) | TSQA (Temporal) | CQA (Complex) | **Average Overall** |
| :--- | :---: | :---: | :---: | :---: |
| **Accuracy (%)** | 45.2 | 38.7 | 31.4 | **38.4** |
| **CIDEr Score** | 0.82 | 0.55 | 0.41 | **0.59** |

*Detailed results and challenge rankings can be found on the [Official DCASE 2025 Results Page](https://dcase.community/challenge2025/task-audio-question-answering-results).*

### 📊 Benchmark Results

Researchers are encouraged to report their Development Set results for comparison against the official DCASE 2025 baselines and top-performing models.

#### Baseline & SOTA Comparison (Dev Set Accuracy %)
The following table compares the performance across the three task subsets. Note that **Part 1** refers to Bioacoustics, **Part 2** to Temporal Soundscapes, and **Part 3** to Complex QA.

| Model | Part 1 (BQA) | Part 2 (TSQA) | Part 3 (CQA) | **Overall Avg** |
| :--- | :---: | :---: | :---: | :---: |
| **Qwen-Omni-2.5 (Chen_SRCN GRPO)** | 66.45 %	| 74.52 %	| 86.05 %|  	81.26 %
| **Gemini-2.0-Flash (Baseline)** | 42.0% | 46.3% | 56.6% | **52.5%** |
| **AudioFlamingo 2 (Baseline)** | 53.9% | 31.7% | 49.5% | **45.7%** |
| **Qwen2-Audio-7B (Baseline)** | 30.0% | 39.2% | 49.6% | **45.0%** |

#### Official Evaluation Leaderboard (Top 3 Snippet)
| Rank | Submission Code | Domain Avg (Eval) | Domain Avg (Dev) |
| :--- | :--- | :---: | :---: |
| 1 | Sun_Antgroup_task5_2 | 73.74% | 77.93% |
| 2 | Shi_USTC_task5_1 | 72.81% | 78.13% |
| 3 | Chen_SRCN_task5_3 | 64.91% | 69.82% |

> **Note:** Baseline results are typically evaluated in a zero-shot setting. For detailed system descriptions and full rankings, please visit the [DCASE 2025 Results Page](https://dcase.community/challenge2025/task-audio-question-answering-results).

---

### Preparing the Multiple Domain Audio (MD-Audio) Training and Dev Data as DCASE 2025 result comparison

```bash
git clone https://huggingface.co/datasets/PeacefulData/2025_DCASE_AudioQA_Official # clone the questions
cd 2025_DCASE_AudioQA_Official
bash download_dcase_25_task5_challenge_audio.sh # download the audio data part 1 will access via Watkins Marine Mammal Sound Database's official link
```

#### Reference
```
@article{yang2025multi,
  title={Multi-domain audio question answering toward acoustic content reasoning in the dcase 2025 challenge},
  author={Yang, Chao-Han Huck and Ghosh, Sreyan and Wang, Qing and Kim, Jaeyeon and Hong, Hengyi and Kumar, Sonal and Zhong, Guirui and Kong, Zhifeng and Sakshi, S and Lokegaonkar, Vaibhavi and others},
  journal={arXiv preprint arXiv:2505.07365},
  year={2025}
}
```