slprl/StressPresso

---
language:
- en
license: cc-by-nc-4.0
task_categories:
- question-answering
- automatic-speech-recognition
- audio-classification
- audio-text-to-text
dataset_info:
features:
- name: transcription
dtype: string
- name: intonation
dtype: string
- name: description
dtype: string
- name: possible_answers
sequence: string
- name: label
dtype: int64
- name: audio_lm_prompt
dtype: string
- name: audio
struct:
- name: array
sequence: float64
- name: path
dtype: string
- name: sampling_rate
dtype: int64
- name: stress_pattern
struct:
- name: binary
sequence: int64
- name: indices
sequence: int64
- name: words
sequence: string
- name: metadata
struct:
- name: audio_path
dtype: string
- name: gender
dtype: string
- name: speaker_id
dtype: string
- name: interpretation_id
dtype: string
- name: transcription_id
dtype: string
splits:
- name: test
num_bytes: 216570205
num_examples: 202
download_size: 135868258
dataset_size: 216570205
tags:
- speech
- stress
- intonation
- audio-reasoning
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
pretty_name: StressPresso
---
# StressPresso Evaluation Dataset

This dataset is derived from the *Expresso* dataset as introduced in the paper
**[EXPRESSO: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis](https://arxiv.org/pdf/2308.05725)**.

For additional information on *Expresso*, see its [project page](https://speechbot.github.io/expresso/).

The *StressPresso* 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?](https://huggingface.co/papers/2505.22765)**

💻 [Code Repository](https://github.com/slp-rl/StressTest) | 🤗 [Model: StresSLM](https://huggingface.co/slprl/StresSLM) | 🤗 [Stress-17k Dataset](https://huggingface.co/datasets/slprl/Stress-17K-raw)

📃 [Paper](https://huggingface.co/papers/2505.22765) | 🌐 [Project Page](https://pages.cs.huji.ac.il/adiyoss-lab/stresstest/)

---

## 🗂️ Dataset Overview

The *StressPresso* dataset includes **202** 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 48kHz sampling rate.
* `metadata`: Structured metadata including:

* `gender`: Speaker gender.
* `audio_path`: Expresso sample name.
* `speaker_id`: Expresso speaker id.
* `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](https://github.com/slp-rl/StressTest).

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

```bash
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 usage with the HuggingFace Datasets library:

```python
from datasets import load_dataset

dataset = load_dataset("slprl/StressPresso")
```

---

## 📖 Citation

If you use this dataset in your work, please cite:

```bibtex
@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},
}
```