metadata

library_name: transformers
language:
  - en
base_model:
  - Qwen/Qwen2-Audio-7B-Instruct
pipeline_tag: audio-text-to-text
tags:
  - lora
license: cc-by-nc-4.0

StresSLM

StresSLM is an audio-text-to-text model fine-tuned with LoRA adapters on top of the Qwen/Qwen2-Audio-7B-Instruct base model. It is designed to tackle Sentence Stress Detection (SSD) and Sentence Stress Reasoning (SSR) tasks on the StressTest benchmark. StresSLM predicts stress patterns and reasoning based on spoken audio. For more information, see our paper and code:

💻 Code | 🤗 StressTest Dataset | 🤗 Stress-17k Dataset

📃 StressTest Paper | 🌐 Project Page

Usage

This model can be loaded using the HuggingFace Transformers library:

from transformers import AutoProcessor, Qwen2AudioForConditionalGeneration
from peft import PeftModel, PeftConfig

# Load processor
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")

# Load LoRA config and base model
peft_config = PeftConfig.from_pretrained("slprl/StresSLM")
base_model = Qwen2AudioForConditionalGeneration.from_pretrained(peft_config.base_model_name_or_path)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "slprl/StresSLM")

Tasks

Sentence Stress Detection (SSD): Identify stressed words in an utterance.
Sentence Stress Reasoning (SSR): Reason about the speaker’s intention using stress patterns.

For evaluation scripts and benchmarks, refer to the StressTest GitHub repository.

📖 Citation

If you use this model, please cite:

@misc{yosha2025stresstest,
      title={StressTest: Can YOUR Speech LM Handle the Stress?},
      author={Iddo Yosha and Gallil Maimon and Yossi Adi},
      year={2025},
      eprint={2505.22765},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.22765},
}