arvinsingh/welsh-speech-landmarks

Welsh Speech Dataset - Facial Landmarks

68-point facial landmarks (ibug68 template) from the Welsh Speech Dataset.

Contents

• Facial landmarks for every frame
• 68 3D points per frame (x, y, z coordinates)
• Format: Parquet
• Manual annotation using ibug68 template

Format

The landmarks.parquet file contains:

Column	Description
speaker_id	Speaker identifier (1-33)
phrase_id	Phrase identifier (1-10)
frame_id	Frame identifier (e.g., "001", "002")
landmarks_json	JSON string containing landmark data

Landmark JSON Structure

{
  "landmarks": {
    "points": [[x1, y1, z1], [x2, y2, z2], ..., [x68, y68, z68]],
    "connectivity": [[0,1], [1,2], ...],
  },
  "labels": [
    {"label": "chin", "mask": [0,1,2,...,16]},
    {"label": "leye", "mask": [17,18,...,22]},
    {"label": "reye", "mask": [23,24,...,28]},
    ...
  ],
  "version": 2
}

Semantic Regions

Region	Points
Chin	0-16
Left Eye	17-22
Right Eye	23-28
Left Eyebrow	29-33
Nose	34-42
Right Eyebrow	43-47
Mouth	48-67

Preview

Landmark Annotations

Joining with Main Metadata

The main metadata (welsh-speech-dataset) is sequence-level (one row per speaker-phrase). Landmarks are frame-level (one row per frame). To get sequence attributes like fluency scores:

import pandas as pd
from huggingface_hub import hf_hub_download

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

# load main metadata (sequence-level)
main_meta_path = hf_hub_download(
    repo_id="arvinsingh/welsh-speech-dataset",
    filename="metadata.parquet",
    repo_type="dataset"
)
main_meta = pd.read_parquet(main_meta_path)

# join on speaker_id and phrase_id to get fluency scores, welsh text, etc.
merged = landmarks.merge(
    main_meta[['speaker_id', 'phrase_id', 'fluency_score', 'welsh_text', 'english_translation']],
    on=['speaker_id', 'phrase_id'],
    how='left'
)

print(merged.head())
# Now each frame has: speaker_id, phrase_id, frame_id, landmarks_json, fluency_score, welsh_text, ...

Basic Usage

import pandas as pd
import json

# load
landmarks = pd.read_parquet("landmarks.parquet")

# get landmarks for specific frame
frame = landmarks[(landmarks['speaker_id'] == 1) &                   (landmarks['phrase_id'] == 1) &                   (landmarks['frame_id'] == '001')].iloc[0]

# parse JSON
landmark_data = json.loads(frame['landmarks_json'])
points = landmark_data['landmarks']['points']  # 68 x 3 array

print(f"Landmark points: {len(points)} points")
print(f"First point (x,y,z): {points[0]}")

Related Repositories

• Main: arvinsingh/welsh-speech-dataset
• Audio: arvinsingh/welsh-speech-audio
• 3D Meshes: arvinsingh/welsh-speech-3d-meshes

Citation

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

@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},
}

@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"},
  year         = 2025,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15397513},
}