---
title: "TutlAit v1: a crowdsourced Moroccan Tamazight — Chadi 2026"
description: "Moroccan Tamazight is an official language spoken by about a quarter of the population."
authors: ["Mohamed-Amine Chadi", "Ezzahra Ait El Arbi", "Ismail Khayoub", "Aymane Fadili", "Yassine Ennhili", "Jadjigua Bouali", "Hanane Inhid", "Mohammed Ameksa", "Hajar Mousannif"]
year: 2026
venue: "arXiv:2609.38219"
source: "https://arxiv.org/abs/2609.38219"
doi: "10.48550/arxiv.2609.38219"
arxiv: "2609.38219"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/chadi-2026-tutlait-v1-crowdsourced-moroccan-tamazight/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/chadi-2026-tutlait-v1-crowdsourced-moroccan-tamazight/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/) / TutlAit v1: a crowdsourced Moroccan Tamazight speech dataset with Arabic transcriptions and regional accent labels

# TutlAit v1: a crowdsourced Moroccan Tamazight speech dataset with Arabic transcriptions and regional accent labels

**Authors:** Mohamed-Amine Chadi, Ezzahra Ait El Arbi, Ismail Khayoub, Aymane Fadili, Yassine Ennhili, Jadjigua Bouali, Hanane Inhid, Mohammed Ameksa, [Hajar Mousannif](https://md.research.elamur.ai/free/en/authors/hajar-mousannif/) · **Year:** 2026 · **Venue:** arXiv:2609.38219 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2609.38219) · [PDF](https://arxiv.org/pdf/2609.38219) · [DOI 10.48550/arxiv.2609.38219](https://doi.org/10.48550/arxiv.2609.38219) · [arXiv 2609.38219](https://arxiv.org/abs/2609.38219) · [OpenAlex](https://openalex.org/W7215079837)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
Moroccan Tamazight is an official language spoken by about a quarter of the population. Yet it has very little public labelled speech, and existing sets do not record which regional variety is spoken. The authors built TutlAit, a crowdsourcing web app (React 18, Django 5 / Django REST Framework, PostgreSQL), to address this gap. Native speakers were recruited on LinkedIn and Instagram, and each declared a regional variety when signing up. They then either recorded a spoken Tamazight version of a displayed Modern Standard Arabic sentence or typed an Arabic rendering of a Tamazight clip. Extra segments came from public media such as YouTube films and series. These were cut and annotated in ELAN and loaded through a bulk CSV/ZIP importer. Every file passes the same checks before release: conversion to 16 kHz mono WAV, SHA-256 duplicate rejection, duration limits, a decoding test and administrator approval. The released corpus has about 20.9 hours of validated speech, each clip carrying a variety label. It can support Tamazight ASR with Arabic-script output, speech translation into Arabic and accent identification.

## Key points
- The corpus has 13,384 WAV files (16 kHz, mono, 16-bit PCM) totalling 75,231 s (≈20.9 h, ≈3.01 GB). Clips average 5.62 s (median 4.03 s), and 19 of them run longer than 30 s.
- The varieties are heavily unbalanced. Atlas has 9,956 files (14.08 h) and Souss has 3,378 files (6.75 h), while Rif (22 files) and Kabyle (28 files) are too small to use alone.
- It is published on Hugging Face as Ma-OpenHub/Tutlait-v1. Each audio file has a row in `metadata.csv` with text, accent, workflow mode, duration, a hashed speaker id, gender, age group, region, sha256 and status. Roughly 5 more hours are still being collected.
- The text side is Modern Standard Arabic, which is effectively a translation of the speech rather than a Tifinagh or Latin transcription. The ELAN-annotated media segments also carry a Latin-script Tamazight tier.
- Accent labels and demographics are self-reported and were not checked by a linguist. Recordings were made on the volunteers' own devices, so noise and microphone quality vary.

## Interesting ideas
- The variety is chosen once, at account creation, and every later contribution inherits it, so each item gets a dialect label at no extra annotation cost. — For Mova (tree): Treebank sentences could carry a source or dialect tag inherited from their corpus, which would let parser accuracy (LAS) be reported separately for each UD dialect conversion.
- Each file is hashed when uploaded so exact duplicates are refused, and the hash is computed again at export to confirm integrity. — For Mova (learn): Normalised sentences in silver data could be hashed at intake to drop repeats and to make sure no evaluation sentence slips into training.
- An optional Gemini helper can pre-fill a translation, but nothing enters the corpus until a human confirms it and a reviewer later approves it. — For Mova (learn): LLM teacher output could be stored as "pending" and promoted to training data only once the witness parsers agree with it.
- Deletions only set a flag (`is_exist`), and the release selects items that are both validated and not flagged, so the audit trail is kept while the published data stays clean. — For Mova (learn): Self-induced repair rules could be retired with a flag instead of being erased, which keeps a history of why each rule was dropped and allows it to be restored.
- The same platform builds pairs in both directions, text prompting speech and speech prompting text, giving two kinds of parallel data from one pool of people. — For Mova (back): The text → graph → text compressor could be tested from both ends, starting from text in one run and from a hand-built meaning graph in the other, to catch errors that only show up in one direction.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{chadi2026tutlait,
  title = {{TutlAit v1: a crowdsourced Moroccan Tamazight speech dataset with Arabic transcriptions and regional accent labels}},
  author = {Chadi, Mohamed-Amine and Arbi, Ezzahra Ait El and Khayoub, Ismail and Fadili, Aymane and Ennhili, Yassine and Bouali, Jadjigua and Inhid, Hanane and Ameksa, Mohammed and Mousannif, Hajar},
  year = {2026},
  doi = {10.48550/arxiv.2609.38219},
  eprint = {2609.38219},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.38219}
}
```

- [← previous Tug-of-war between idioms' figurative and literal interpretations in LLMs](https://md.research.elamur.ai/free/en/papers/oh-2026-tug-war-between-idioms-figurative/)
- [next → Ugaritic Parallel of Jahwe melek ʿōlām](https://md.research.elamur.ai/free/en/papers/wozniak-2026-ugaritic-parallel-jahwe-melek-olam/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/)

## References

13 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Royaume du Maroc: Constitution du Royaume du Maroc. Bulletin Officiel. Article 5: recognition of Tamazight as an official language (2011). [https://www.ircam.ma/fr/textes-fondateurs/constitution-du-royaume](https://www.ircam.ma/fr/textes-fondateurs/constitution-du-royaume) [↗](https://www.ircam.ma/fr/textes-fondateurs/constitution-du-royaume)
2. Haut-Commissariat au Plan (HCP): Trois notes explicatives relatives au RGPH 2024. [https://www.hcp.ma/Trois-notes-explicatives-relatives-au-RGPH-2024](https://www.hcp.ma/Trois-notes-explicatives-relatives-au-RGPH-2024) a4039.html. Accessed 19 June 2026 (2024) [↗](https://www.hcp.ma/Trois-notes-explicatives-relatives-au-RGPH-2024)
3. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: Proceedings of the International Conference on Machine Learning (ICML), pp. 28492–28518 (2023). https://arxiv.org/abs/2212.04356 [↗](https://arxiv.org/abs/2212.04356) [our card](https://md.research.elamur.ai/free/en/papers/radford-2022-robust-speech-recognition-large-scale/)
4. Öktem, A., Boudichat, L.: Awal – Community-Powered Language Technology for Tamazight. arXiv preprint arXiv:2510.27407 (2025). https://arxiv.org/abs/2510. 27407 [↗](https://arxiv.org/abs/2510.27407)
5. Dhimi, S.: Tamazight ASR Dataset v2. Hugging Face Datasets (2024). [https://huggingface.co/datasets/SoufianeDahimi/Tamazight-ASR-Dataset-v2](https://huggingface.co/datasets/SoufianeDahimi/Tamazight-ASR-Dataset-v2) [↗](https://huggingface.co/datasets/SoufianeDahimi/Tamazight-ASR-Dataset-v2)
6. Akallouch, M., et al. : Advances in Amazigh language technologies: A comprehensive survey. Information 16 (7), 600 (2025) [find](https://scholar.google.com/scholar?q=Akallouch%2C+M%2E%2C+et+al%2E+%3A+Advances+in+Amazigh+language+technologies%3A+A+comprehensive+survey%2E+Information+16+%287%29%2C+600+%282025%29)
7. Chadi, M.-A., Ait El Arbi, E., Khayoub, I., Fadili, A., Ennhili, Y., Bouali, J., Inhid, H., Ameksa, M., Mousannif, H.: TutlAit v1: A crowdsourced Moroccan Tamazight speech dataset with Arabic transcriptions and regional accent labels. Hugging Face Datasets (2026). [https://huggingface.co/datasets/Ma-OpenHub/](https://huggingface.co/datasets/Ma-OpenHub/) Tutlait-v1 [↗](https://huggingface.co/datasets/Ma-OpenHub/) [our card](https://md.research.elamur.ai/free/en/papers/chadi-2026-tutlait-v1-crowdsourced-moroccan-tamazight/)
8. Meta Platforms, Inc.: React: A JavaScript Library for Building User Interfaces. [https://react.dev](https://react.dev) (2013) [↗](https://react.dev)
9. Jones, M., Bradley, J., Sakimura, N.: JSON Web Token (JWT). RFC 7519, Internet Engineering Task Force (IETF) (2015) [find](https://scholar.google.com/scholar?q=Jones%2C+M%2E%2C+Bradley%2C+J%2E%2C+Sakimura%2C+N%2E%3A+JSON+Web+Token+%28JWT%29%2E+RFC+7519%2C+Internet+Engineering+Task+Force+%28IETF%29+%282015%29)
10. Christie, T., et al.: Django REST django-rest-framework.org. Version 3.x (2011) Framework. [https://www](https://www). [↗](https://www)
11. The PostgreSQL Global Development Group: PostgreSQL: The World’s Most Advanced Open Source Relational Database. [https://www.postgresql.org](https://www.postgresql.org) (2023) [↗](https://www.postgresql.org)
12. Max Planck Institute for Psycholinguistics, The Language Archive: ELAN (Version 6.8). Software (2024). [https://archive.mpi.nl/tla/elan](https://archive.mpi.nl/tla/elan) [↗](https://archive.mpi.nl/tla/elan)
13. National Institute of Standards and Technology: Secure Hash Standard (SHS). FIPS Publication 180-4 (2015) [find](https://scholar.google.com/scholar?q=National+Institute+of+Standards+and+Technology%3A+Secure+Hash+Standard+%28SHS%29%2E+FIPS+Publication+180%2D4+%282015%29)