---
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 22050
  - name: speaker_id
    dtype: string
  - name: sentence_id
    dtype: string
  - name: duration
    dtype: float64
  - name: text
    dtype: string
  - name: split
    dtype: string
  splits:
  - name: train
    num_bytes: 1573717458
    num_examples: 2373
  - name: test
    num_bytes: 1119588307
    num_examples: 1510
  - name: validation
    num_bytes: 213986775
    num_examples: 268
  - name: held_out
    num_bytes: 531217549
    num_examples: 720
  download_size: 2799400227
  dataset_size: 3438510089
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
  - split: held_out
    path: data/held_out-*
authors:
- lenaso
pretty_name: Saar-Voice
---

# Saar-Voice

A multi-speaker speech corpus for the Rhine Franconian dialect of German as spoken in Saarbrücken and the surrounding region, referred to loosely as **Saarländisch**.

## Overview

| Property | Value |
|---|---|
| Speakers | 9 (P01–P09) |
| Total duration | ~6 hours |
| Sentences (recorded) | 4,871 |
| Sentences (unrecorded) | 3,901 |
| Sampling rate | 22,050 Hz |
| Language | Saarländisch (Rhine Franconian) |
| Format | WAV (stereo, 16-bit) |

## Splits

| Split | Sentences | Purpose |
|---|---|---|
| `train` | 2,373 | Model training |
| `test` | 1,510 | Test set |
| `validation` | 268 | Validation set |
| `held_out` | 720 | Held-out speaker evaluation |

The `held_out` split contains recordings from speakers not seen during training, intended for further tests such as transfer learning experiments.

The dataset additionally includes [`unrecorded.json`](./unrecorded.json), a list of 3,905 sentences for which no audio was recorded, provided for completeness and for future work.

## Speakers

The corpus includes 9 speakers (P01–P09) of varying age and gender, all native speakers of Saarländisch. Speaker metadata (age, gender) is not included in this release.

## Text

Sentences were sourced from:

- the [MASSIVE dataset](https://huggingface.co/datasets/AmazonScience/massive) (German subset) + localized into Saarländisch dialect orthography
- several books, poem and short story collections
- texts provided directly by the local author community

A detailed breakdown is available in the corresponding paper.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("UdS-LSV/Saar-Voice")

# Access the training split
for example in ds["train"]:
    audio = example["audio"]          # dict with array, sampling_rate, path
    text  = example["text"]           # dialect transcription
    spk   = example["speaker_id"]     # e.g. "P01"
```

## Citation

If you use this dataset, please cite:

```bibtex
@misc{oberkircher2026saarvoice,
  title  = {Saar-Voice: A Multi-Speaker Saarbrücken Dialect Speech Corpus},
  author = {Lena S. Oberkircher and Jesujoba O. Alabi and Dietrich Klakow and Jürgen Trouvain},
  year   = {2026},
  url    = {https://arxiv.org/abs/2604.11803}
}
```

## License

This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

## Acknowledgements