README.md

---
license: cc-by-nc-4.0
task_categories:
  - automatic-speech-recognition
  - text-to-speech
  - text-to-audio
language:
  - cy
tags:
  - speech
  - welsh
  - cymraeg
  - 3d-face
  - facial-landmarks
  - multimodal
  - fluency
  - pronunciation
  - 4d-dataset
size_categories:
  - 100K<n<1M
pretty_name: CymruFluency Welsh Speech Dataset
---

# Welsh Speech Dataset

A multimodal dataset of 33 speakers producing 10 Welsh phrases, captured using 3DMD technology with audio and dense facial landmark annotations.

## Dataset Overview

- **Speakers**: 33 participants
- **Phrases**: 10 Welsh phrases per speaker
- **Sequences**: ~330 (33 speakers x 10 phrases)
- **Modalities**: 
  - Audio recordings (.wav)
  - 3D facial reconstructions (.obj meshes + texture maps)
  - 68-point facial landmarks (ibug68 template)
- **Fluency Scores**: Each phrase rated 0-5 (5 = perfect, 0 = many errors)

## Preview

<div align="center">
  <video style="object-fit: cover;" controls loop src="https://github.com/user-attachments/assets/140c9079-a195-4c05-aec0-7b9876048030" muted="false"></video>
  <p><strong>Subject uttering Welsh phrase “Gwybodaeth angenrheidiol” (Tr. EN: Necessary information; IPA: /ˈɡʊɨ̯bɔðaɪθ aŋɛnˈhreɪ̯djɔl/)</strong></p>
</div>

## Repository Structure

This dataset is split into 4 repositories for convenience:

1. **[welsh-speech-dataset](https://huggingface.co/datasets/arvinsingh/welsh-speech-dataset)** (this repo) - Main hub with sequence-level metadata
2. **[welsh-speech-audio](https://huggingface.co/datasets/arvinsingh/welsh-speech-audio)** - Audio recordings only
3. **[welsh-speech-3d-meshes](https://huggingface.co/datasets/arvinsingh/welsh-speech-3d-meshes)** - 3D facial meshes (zipped per sequence)
4. **[welsh-speech-landmarks](https://huggingface.co/datasets/arvinsingh/welsh-speech-landmarks)** - Facial landmarks (frame-level Parquet)

## Metadata

The `metadata.csv` and `metadata.parquet` files contain **sequence-level** data (one row per speaker-phrase):

| Column | Description |
|--------|-------------|
| `speaker_id` | Speaker identifier (1-33) |
| `phrase_id` | Phrase identifier (1-10) |
| `audio_path` | Path to audio file |
| `mesh_zip_path` | Path to 3D mesh zip file |
| `fluency_score` | Pronunciation quality score (0-5) |
| `welsh_text` | Welsh phrase text |
| `english_translation` | English translation |
| `num_frames` | Number of frames in the sequence |
| `has_3d` | Boolean indicating 3D data availability |
| `has_landmark` | Boolean indicating landmark availability |

**Note:** Landmark data is stored in `landmarks.parquet` in the landmarks repository at **frame-level**. 
Join using `speaker_id` and `phrase_id` to combine with this sequence-level metadata.

## Welsh Phrases

| ID | Welsh Text | English Translation |
|----|-----------|---------------------|
| 1 | Eisteddfod yr Urdd | Welsh Youth Music Competition |
| 2 | Prynhawn da bawb | Good afternoon everyone |
| 3 | Dyn busnes yw e | It's a businessman |
| 4 | Papur a phensil | Paper and pencil |
| 5 | Ardderchog | Excellent / Superb |
| 6 | Llwyddiant ysgubol | Great success |
| 7 | Yng nghanol y dref | In the town center |
| 8 | Dwy neuadd gymunedol | Two community halls |
| 9 | Llunio rhestr fer | Shortlisted |
| 10 | Gwobodaeth angenrheidiol | Necessary information |

## Usage

### Load Metadata

```python
import pandas as pd

# load sequence-level metadata
metadata = pd.read_parquet("metadata.parquet")

# filter by fluency score
high_quality = metadata[metadata['fluency_score'] >= 4]

# get info for specific speaker/phrase
seq = metadata[(metadata['speaker_id'] == 1) & (metadata['phrase_id'] == 1)].iloc[0]
print(f"Frames: {seq['num_frames']}, Fluency: {seq['fluency_score']}")
```

### Access Specific Modalities

Download only what you need:

```python
from huggingface_hub import hf_hub_download
import zipfile

# download audio
audio_file = hf_hub_download(
    repo_id="arvinsingh/welsh-speech-audio",
    filename="audio/speaker_01_phrase_01.wav",
    repo_type="dataset"
)

# download 3D mesh zip for a sequence
mesh_zip = hf_hub_download(
    repo_id="arvinsingh/welsh-speech-3d-meshes",
    filename="meshes/speaker_01_phrase_01.zip",
    repo_type="dataset"
)

# extract meshes
with zipfile.ZipFile(mesh_zip, 'r') as zf:
    zf.extractall("speaker_01_phrase_01")
    # Contains: 001.obj, 001.png, 002.obj, 002.png, ...

# load landmarks (frame-level)
import pandas as pd
landmarks = pd.read_parquet(
    hf_hub_download(
        repo_id="arvinsingh/welsh-speech-landmarks",
        filename="landmarks.parquet",
        repo_type="dataset"
    )
)

# join landmarks with main metadata to get fluency scores
merged = landmarks.merge(metadata, on=['speaker_id', 'phrase_id'])
```

## Citation

If you use this dataset, please cite both the paper and the dataset:

```bibtex
@inproceedings{bali_2026_cymrufluency,
  author       = {Bali, Arvinder Pal Singh and
                  Tam, Gary KL and
                  Siris, Avishek and
                  Andrews, Gareth and
                  Lai, Yukun and
                  Tiddeman, Bernie and
                  Ffrancon, Gwenno},
  title        = {CymruFluency - A Fusion Technique and a 4D Welsh Dataset for Welsh Fluency Analysis},
  booktitle    = {Advanced Concepts for Intelligent Vision Systems},
  pages        = {96--108},
  year         = 2026,
  publisher    = {Springer Nature Switzerland},
  doi          = {10.1007/978-3-032-07343-3_8},
  url          = {https://doi.org/10.1007/978-3-032-07343-3_8},
}

@dataset{bali_2025_dataset,
  author       = {Bali, Arvinder Pal Singh and
                  Tam, Gary KL and
                  Siris, Avishek and
                  Andrews, Gareth and
                  Lai, Yukun and
                  Tiddeman, Bernie and
                  Ffrancon, Gwenno},
  title        = {Dataset and code for "CymruFluency - A fusion technique and a 4D Welsh dataset for Welsh fluency analysis"},
  month        = may,
  year         = 2025,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15397513},
  url          = {https://doi.org/10.5281/zenodo.15397513},
}
```

## Original Data

The original dataset is published on Zenodo: [10.5281/zenodo.15397513](https://doi.org/10.5281/zenodo.15397513)

## License

Creative Commons Attribution-NonCommercial 4.0 International License.

## Acknowledgments

Dataset collected using 3DMD facial capture technology. All frames manually annotated with ibug68 facial landmarks.