README.md

metadata

dataset_info:
  features:
    - name: transcription_id
      dtype: string
    - name: transcription
      dtype: string
    - name: description
      dtype: string
    - name: intonation
      dtype: string
    - name: interpretation_id
      dtype: string
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: metadata
      struct:
        - name: gender
          dtype: string
        - name: language_code
          dtype: string
        - name: sample_rate_hertz
          dtype: int64
        - name: voice_name
          dtype: string
    - name: possible_answers
      sequence: string
    - name: label
      dtype: int64
    - name: stress_pattern
      struct:
        - name: binary
          sequence: int64
        - name: indices
          sequence: int64
        - name: words
          sequence: string
    - name: audio_lm_prompt
      dtype: string
  splits:
    - name: test
      num_bytes: 29451897.32142857
      num_examples: 218
  download_size: 22754357
  dataset_size: 29451897.32142857
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
license: cc-by-nc-4.0

StressTest Evaluation Dataset

This dataset supports the evaluation of models on Sentence Stress Reasoning (SSR) and Sentence Stress Detection (SSD) tasks, as introduced in our paper:

StressTest: Can YOUR Speech LM Handle the Stress?

📃 Paper | 💻 Code Repository | 🤗 Model: StresSLM

---

🗂️ Dataset Overview

This dataset includes 218 evaluation samples (split: test) with the following features:

• transcription_id: Identifier for each transcription sample.

• transcription: The spoken text.

• description: Description of the interpretation of the stress pattern.

• intonation: The stressed version of the transcription.

• interpretation_id: Unique reference to the interpretation imposed by the stress pattern of the sentence.

• audio: Audio data at 16kHz sampling rate.

• metadata: Structured metadata including:

  • gender: Speaker gender.
  • language_code: Language of the transcription.
  • sample_rate_hertz: Sampling rate in Hz.
  • voice_name: Voice name.

• possible_answers: List of possible interpretations for SSR.

• label: Ground truth label for SSR.

• stress_pattern: Structured stress annotation including:

  • binary: Sequence of 0/1 labels marking stressed words.
  • indices: Stressed word positions in the transcription.
  • words: The actual stressed words.

• audio_lm_prompt: The prompt used for SSR.

---

Evaluate YOUR model

This dataset is designed for evaluating models following the protocol and scripts in our StressTest repository.

To evaluate a model, refer to the instructions in the repository. For example:

python -m stresstest.evaluation.main \
    --task ssr \
    --model_to_evaluate stresslm

Replace ssr with ssd for stress detection, and use your model’s name with --model_to_evaluate.

---

How to use

This dataset is formatted for with the HuggingFace Datasets library:

from datasets import load_dataset

dataset = load_dataset("slprl/StressTest")

---

📖 Citation

If you use this dataset in your work, 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},
}