TutlAit
A Moroccan Tamazight speech dataset that pairs Tamazight speech with Modern Standard Arabic text and includes regional accent labels. Collected via crowdsourcing, it is intended to support automatic speech recognition and speech translation research for low-resource languages.
Timeline (last 90 days)
TutlAit is a corpus that pairs Moroccan Tamazight speech with Modern Standard Arabic text and carries explicit regional accent labels.
The TutlAit crowdsourcing web application uses React 18 for the frontend, Django 5 and Django REST Framework for the backend, and PostgreSQL for data storage.
The research team recruited native speakers through targeted promotional campaigns on LinkedIn and Instagram.
TutlAit adopts two workflows: Text-to-Audio and Audio-to-Text.
The team used the ELAN tool to segment and annotate supplementary clips extracted from audiovisual media, then imported them into the system via a batch CSV/ZIP pipeline.
TutlAit uniformly converts audio to 16 kHz mono WAV format on the server side and uses SHA-256 hashing for deduplication.
TutlAit's data quality control combines hash-based deduplication, duration boundary checks, and manual review.
The TutlAit v1 dataset contains 13,384 audio files with a total duration of 75,231 seconds (approximately 20.9 hours, approximately 3.01 GB).
The regional variety distribution in TutlAit v1 is uneven: the Atlas variety has 9,956 files (approximately 14.08 hours), the Souss variety has 3,378 files (approximately 6.75 hours), the Rif variety has only 22 files, and the Kabyle variety has only 28 files.
The Rif and Kabyle variety subsets are extremely small in scale, which is a significant shortcoming for training models that cover all varieties.
All Facts (10)
TutlAit is a corpus that pairs Moroccan Tamazight speech with Modern Standard Arabic text and carries explicit regional accent labels.
50%UnverifiedThe TutlAit crowdsourcing web application uses React 18 for the frontend, Django 5 and Django REST Framework for the backend, and PostgreSQL for data storage.
50%UnverifiedThe research team recruited native speakers through targeted promotional campaigns on LinkedIn and Instagram.
50%UnverifiedTutlAit adopts two workflows: Text-to-Audio and Audio-to-Text.
50%UnverifiedThe team used the ELAN tool to segment and annotate supplementary clips extracted from audiovisual media, then imported them into the system via a batch CSV/ZIP pipeline.
50%UnverifiedTutlAit uniformly converts audio to 16 kHz mono WAV format on the server side and uses SHA-256 hashing for deduplication.
50%UnverifiedTutlAit's data quality control combines hash-based deduplication, duration boundary checks, and manual review.
50%UnverifiedThe TutlAit v1 dataset contains 13,384 audio files with a total duration of 75,231 seconds (approximately 20.9 hours, approximately 3.01 GB).
50%UnverifiedThe regional variety distribution in TutlAit v1 is uneven: the Atlas variety has 9,956 files (approximately 14.08 hours), the Souss variety has 3,378 files (approximately 6.75 hours), the Rif variety has only 22 files, and the Kabyle variety has only 28 files.
50%UnverifiedThe Rif and Kabyle variety subsets are extremely small in scale, which is a significant shortcoming for training models that cover all varieties.
50%