📑 Paper | 👩‍💻 Github | 🤗 Model | 🤗 Dataset

DeSTA2.5-Audio: Toward General-Purpose Large Audio Language Model with Self-Generated Cross-Modal Alignment

Self-generated data is what you need for developing general-purpose LALMs!

• 🧪 A new training framework (read the paper)
  • Highly scalable and efficient without task-specific instruction-tuning data
  • Preserves language ability and avoids catastrophic forgetting
  • Comprehensive studies on data quality in LALM development
• 📦 Open resources for the community
  • Model checkpoints and Training scripts
  • DeSTA-AQA5M dataset (5M audio-text pairs from 7,000 hours of audio)

✨ News / Change logs

• 🚧 Coming soon: vLLM-based data construction script, detailed finetuning tutorials
• 2025/07/23: Released training scripts. Now you can train your own DeSTA-Audio. 📘 Training README
• 2025/07/21: Released DeSTA-AQA-5M dataset! 📘 Dataset README 🤗 DeSTA-AQA5M
• 2025/07/10: Inference code and model checkpoints are live! 🤗 DeSTA2.5-Audio
• 2025/07/03: DeSTA2.5-Audio paper is on arXiv! 📄 arXiv:2507.02768

📄 Documents

🧐 Architecture

🚀Quickstart

Installation

git clone https://github.com/kehanlu/DeSTA2.5-Audio.git
cd DeSTA2.5-Audio
pip install -e .

Basic Usage

from desta import DeSTA25AudioModel

# Load the model from Hugging Face
model = DeSTA25AudioModel.from_pretrained("DeSTA-ntu/DeSTA2.5-Audio-Llama-3.1-8B")
model.to("cuda")

# Run inference with audio input
messages = [
    {
        "role": "system",
        "content": "Focus on the audio clips and instructions."
    },
    {
        "role": "user",
        "content": "<|AUDIO|>\nDescribe this audio.",
        "audios": [{
            "audio": "/path/to/audio.wav",  # Path to your audio file
            "text": None
        }]
    }
]

outputs = model.generate(
    messages=messages,
    do_sample=False,
    top_p=1.0,
    temperature=1.0,
    max_new_tokens=512
)

print(outputs.text)

📂 Dataset

See docs/dataset.md for more details.

🚆 Training & Finetuning

See docs/train.md for more details.

bash example/train/train_example.sh

📚 Citation

@article{lu2025desta25Audio,
  title={DeSTA2.5-Audio: Toward General-Purpose Large Audio Language Model with Self-Generated Cross-Modal Alignment},
  author={Lu, Ke-Han and Chen, Zhehuai and Fu, Szu-Wei and Yang, Chao-Han Huck and Huang, Sung-Feng and Yang, Chih-Kai and Yu, Chee-En and Chen, Chun-Wei and Chen, Wei-Chih and Huang, Chien-yu and others},
  journal={arXiv preprint arXiv:2507.02768},
  year={2025}
}

@inproceedings{lu2025developing,
  title={Developing instruction-following speech language model without speech instruction-tuning data},
  author={Lu, Ke-Han and Chen, Zhehuai and Fu, Szu-Wei and Yang, Chao-Han Huck and Balam, Jagadeesh and Ginsburg, Boris and Wang, Yu-Chiang Frank and Lee, Hung-yi},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

@inproceedings{lu24c_interspeech,
  title     = {DeSTA: Enhancing Speech Language Models through Descriptive Speech-Text Alignment},
  author    = {Ke-Han Lu and Zhehuai Chen and Szu-Wei Fu and He Huang and Boris Ginsburg and Yu-Chiang Frank Wang and Hung-yi Lee},
  year      = {2024},
  booktitle = {Interspeech 2024},
  pages     = {4159--4163},
  doi       = {10.21437/Interspeech.2024-457},
  issn      = {2958-1796},
}

👥 Contributors

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu, Chao-Han Huck Yang, Sung-Feng Huang, Chih-Kai Yang, Chee-En Yu, Chun-Wei Chen, Wei-Chih Chen, Chien-yu Huang, Yi-Cheng Lin, Yu-Xiang Lin, Chi-An Fu, Chun-Yi Kuan, Wenze Ren, Xuanjun Chen, Wei-Ping Huang, En-Pei Hu, Tzu-Quan Lin, Yuan-Kuei Wu, Kuan-Po Huang, Hsiao-Ying Huang, Huang-Cheng Chou, Kai-Wei Chang, Cheng-Han Chiang, Boris Ginsburg, Yu-Chiang Frank Wang, Hung-yi Lee