---
license: cc-by-nc-4.0
task_categories:
- audio-classification
tags:
- audio-forensics
- synthetic-speech-detection
- challenge
- text-to-speech
language:
- en
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: audio
    dtype:
      audio:
        decode: false
  splits:
  - name: test
    num_bytes: 51383393.0
    num_examples: 100
  download_size: 50947225
  dataset_size: 51383393.0
---

# SAFE Challenge Practice Dataset

This repository hosts the practice dataset for the [Audio Forensics Evaluation (SAFE) Challenge](https://huggingface.co/papers/2510.03387). The SAFE Challenge is a fully blind evaluation framework designed to benchmark detection models across progressively harder scenarios: raw synthetic speech, processed audio (e.g., compression, resampling), and laundered audio intended to evade forensic analysis.

The challenge aims to advance the state of the art in audio forensics by driving innovation in detecting and attributing synthetic and manipulated audio artifacts. This practice dataset is provided as a small sample specifically for participants to troubleshoot their model submissions for the competition.

-   **Paper:** [Audio Forensics Evaluation (SAFE) Challenge](https://huggingface.co/papers/2510.03387)
-   **Project Page:** https://stresearch.github.io/SAFE/
-   **Code (Challenge Repository):** https://github.com/stresearch/SAFE

## Dataset Details

This practice dataset is constructed from multiple sources, containing human and machine generated speech audio tracks:
-   https://huggingface.co/SWivid/F5-TTS
-   https://keithito.com/LJ-Speech-Dataset/
-   https://librivox.org

The license of the original sources applies to those respective parts of the dataset.

## Sample Usage

You can easily load this practice dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

# Load the practice dataset
dataset = load_dataset("safe-challenge/safe-challenge-practice-dataset", split="test")

# Access an example
print(dataset[0]["id"])
print(dataset[0]["audio"]["path"])
# Note: By default, the audio is not decoded into an array (due to decode: false in metadata).
# To load the audio as an array, you can cast the 'audio' column. For example:
# from datasets import Audio
# dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
# print(dataset[0]["audio"]["array"])
```

For more details on preparing model submissions and interacting with the challenge, please refer to the [SAFE Challenge GitHub repository](https://github.com/stresearch/SAFE) and the provided [debug example](https://github.com/stresearch/SAFE/blob/main/debug_example.md).