Datasets:

KBayoud
/

ASR_TEDx_Tunisie

license: cc-by-nc-nd-4.0
task_categories:
  - automatic-speech-recognition
language:
  - ar
tags:
  - Tunisian
  - dialect
  - Arabic
  - ASR
  - Speech
  - Translation
pretty_name: TEDxTN
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: transcription
      dtype: string
    - name: duration
      dtype: float64
    - name: start_time
      dtype: float64
    - name: end_time
      dtype: float64
    - name: youtube_video
      dtype: string
  splits:
    - name: train
      num_bytes: 2425153587
      num_examples: 15023
    - name: validation
      num_bytes: 130598416
      num_examples: 731
    - name: test
      num_bytes: 143494811
      num_examples: 840
  download_size: 2739037656
  dataset_size: 2699246814

Dataset Source

Curated by: Fethi Bougares
Language(s) (NLP) : Tunisian Arabic
License: CC BY-NC-ND 4.0 license

BibTeX:

@inproceedings{bougares-etal-2025-tedxtn,
    title = "{TED}x{TN}: A Three-way Speech Translation Corpus for Code-Switched {T}unisian {A}rabic - {E}nglish",
    author = "Bougares, Fethi  and Mdhaffar, Salima  and Elleuch, Haroun  and Est{\`e}ve, Yannick",
    booktitle = "Proceedings of The Third Arabic Natural Language Processing Conference",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.arabicnlp-main.22/",
    doi = "10.18653/v1/2025.arabicnlp-main.22",
    pages = "278--287",
    ISBN = "979-8-89176-352-4"
}

Changes made:

Extracted only transcription segments, skipping the .en translation files.
Added YouTube video URLs for each segment.
Standardized audio segment filenames to match the YouTube video ID and start/end times.
Verified audio segments can be loaded with HuggingFace datasets library.
Saved as and HF Dataset ready for ML pipelines.

Features:

audio: 16 kHz WAV audio segment
transcription: corresponding transcription text
youtube_video: original YouTube URL
start_time, end_time, duration: segment timings """