Dataset Card for SADA - Saudi Audio Dataset for Arabic

Dataset Details

Dataset Description

The SADA (Saudi Audio Dataset for Arabic) is a comprehensive dataset consisting of audio recordings from over 57 TV shows aired by the Saudi Broadcasting Authority (SBA). The dataset contains approximately 667 hours of audio data with transcripts, the majority of which are in various Saudi dialects (Najdi, Hijazi, Khaliji, etc.).

• Curated by: The National Center for Artificial Intelligence at the Saudi Data and Artificial Intelligence Authority (SDAIA) in collaboration with SBA.
• Language(s) (NLP): Arabic (ar)
• License: CC BY-NC-SA 4.0
• Number of audio files: 4563

Dataset Sources

• Repository: Saudi Broadcasting Authority (SBA)

Uses

Direct Use

The dataset is intended for use in automatic speech recognition (ASR), text-to-speech (TTS) systems, and dialect identification tasks. It provides rich linguistic diversity across various dialects, making it suitable for training and fine-tuning speech models for the Arabic language.

Out-of-Scope Use

The dataset should not be used for commercial purposes or any misuse involving speech or text processing in harmful or inappropriate contexts.

Dataset Structure

Features

The dataset contains the following features:

• audio: The audio files in .wav format.
• ProcessedText: The cleaned and pre-processed transcription of the audio segment.
• ShowName: The name of the TV show from which the audio is sourced.
• FullFileLength: Duration of the full audio file in seconds.
• SegmentID: Unique ID for each audio segment.
• SegmentLength: Length of each segment in seconds.
• SegmentStart/End: Start and end times of the segment within the full audio.
• SpeakerAge: Age group of the speaker (e.g., Adult, Child).
• SpeakerGender: Gender of the speaker (Male, Female).
• SpeakerDialect: The dialect spoken by the speaker (e.g., Najdi, Hijazi).
• Environment: The environment in which the audio was recorded (e.g., Clean, Noisy).
• Category: The genre of the show (e.g., مسابقات, درامي, وثائقي).

Dataset Creation

Curation Rationale

The dataset was created to provide a comprehensive resource for developing Arabic speech models, specifically targeting the variety of Saudi dialects and their applications in speech recognition and synthesis.

Source Data

Data Collection and Processing

The data was collected from TV shows aired by the Saudi Broadcasting Authority. The audio files were segmented and transcribed manually, with a focus on ensuring transcription accuracy and uniformity in formatting.

Who are the source data producers?

The original content comes from the Saudi Broadcasting Authority's shows, curated by SDAIA.

Personal and Sensitive Information

The dataset does not contain sensitive personal information, as the focus is on broadcast TV shows with public speakers.

Bias, Risks, and Limitations

Given the focus on Saudi dialects, the dataset may not fully represent other Arabic dialects or languages. Users should be aware that models trained on this dataset may show biases toward Saudi dialects.

Citation

@inproceedings{SADA2023, title={SADA - SBA & SDAIA Audio Dataset for Arabic}, author={Areeb Alowisheq, Abdullah Alrajeh, Sadeen Alharbi, Abdulmajeed Alrowithi, Aljawharah Bin Tamran, Asma Ibrahim, Raghad Aloraini, Raneem Alnajim, Ranya Alkahtani, Renad Almuasaad, Sara Alrasheed, Shaykhah Alsubaie, Yaser Alonaizan}, booktitle={To be published}, affiliation={NCAI-SDAIA}, year={2023} }

Contact

For any inquiries, please contact the National Center for Artificial Intelligence at SDAIA.