Elamur Research⌕
☰
home / papers / 2025 / LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect

LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect

Summary

Tunisian Arabic is hard for speech recognition. It has no standard spelling, pronunciation varies by region, and speakers often switch into French or English. Very little of it is transcribed: TARIC covers a narrow train-station domain, TunSwitch is small, and MASC holds only about 3 hours of Tunisian. Large multilingual systems such as Whisper perform poorly on it.

The authors release three CC BY 4.0 datasets on Hugging Face:

  • a text corpus compiled from public datasets, crawled stories, blogs, transcripts, rap lyrics and tweets;
  • about 93 hours of transcribed audio from TV, radio, podcasts, storytelling and educational sites;
  • an augmented version of that audio, grown to 466 hours through noise separation and voice conversion.

Transcripts were cleaned with a spelling-normalization dictionary and semi-automatic Arabizi transliteration.

On this data they trained a Kaldi TDNN baseline with a 4-gram language model. It clearly outperforms Whisper large models on every test subset. Voice conversion augmentation mainly improves recognition of code-switched Latin-script words. The result is the first openly trained Tunisian ASR baseline, with reusable training and evaluation material.

Key points

  • Text corpus (linto-dataset-text-ar-tn): about 4.5M lines, about 17M words and roughly 289k unique words, including Arabic, French/English and mixed code-switched lines.
  • Spelling normalization: a dictionary maps over 12,500 variant spellings to canonical forms. Arabizi from social media was transliterated by correcting automatic output by hand, because existing MSA-oriented Arabizi tools did not suit Tunisian.
  • Audio datasets:
    • linto-dataset-audio-ar-tn holds about 93 hours at 16 kHz. Most YouTube transcripts were auto-captions, later repaired with the dictionary and manual listening.
    • The augmented set reaches 466 hours. It uses Deezer Spleeter for speech/music separation and SoftVC VITS voice conversion (HuBert encoder, HiFi-GAN decoder) trained on voices from 7 speakers.
  • Whisper results: Whisper large v2/v3/v3 turbo reach WER between about 50% and over 100%, the latter driven by insertions.
  • Kaldi baseline results: on TunSwitch code-switched speech, WER falls from about 58% with no augmentation to about 25% with classical augmentation and about 20% with voice conversion. The voice-conversion model also improves recall on Latin words.

Interesting ideas

  • I1. A single canonical-spelling dictionary is used for two jobs: training the language model and repairing truncated or misheard auto-transcripts, so noisy sources become consistent with the rest of the data. — For Mova (tokens): A variant-to-canonical map for informal English spellings (gonna, tho, cuz) could feed the lemmatizer and also normalize noisy web text before it enters silver data.
  • I2. Augmentation is aimed at a specific weakness: voice conversion was applied mostly to sources dominated by one speaker, rather than spread uniformly over everything. — For Mova (learn): When building silver data, measure which sources or styles are over-represented and generate controlled variants (names, paraphrases) only for those slices to avoid overfitting to one author's style.
  • I3. The minority phenomenon gets its own score: alongside overall WER, the authors report precision, recall and F1 on Latin-script words in code-switched speech, with bootstrap confidence intervals. This exposed gains that the global metric understates. — For Mova (learn): Report separate scores with bootstrap intervals for rare constructions (e.g., gapping, inverted clauses) so a parser change that helps them is not hidden by LAS.
  • I4. Splitting errors into insertions, deletions and substitutions explained WER above 100% as hallucinated insertions rather than ordinary misrecognition. — For Mova (tree): Decomposing parser errors into missing arcs, spurious arcs and wrong labels/heads would show whether a model is guessing structure or mis-attaching it.
  • I5. The authors checked whether tools built for the standard language actually transfer to the dialect, found they did not, and fell back on automatic output plus human correction. — For Mova (tree): Before converting or reusing a UD treebank from another dialect or register, test the converter on a small hand-checked sample and route failures to correction instead of assuming transfer.

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

Cited by

4 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 3 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2504.02604 · arXiv 2504.02604

@misc{naouara2025linto,
  title = {{LinTO Audio and Textual Datasets to Train and Evaluate Automatic Speech Recognition in Tunisian Arabic Dialect}},
  author = {Naouara, Hedi and Lorre, Jean-Pierre and Louradour, Jerome},
  year = {2025},
  doi = {10.48550/arxiv.2504.02604},
  eprint = {2504.02604},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2504.02604}
}

References

23 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abdallah, A. A. B.; Kabboudi, A.; Kanoun, A.; and Zaiem, S. 2023. Leveraging Data Collection and Unsupervised Learning for Code-switched Tunisian Arabic Automatic Speech Recognition. arXiv:2309.11327. ↗ our card
  2. Al-Fetyani, M.; Al-Barham, M.; Abandah, G.; Alsharkawi, A.; and Dawas, M. 2021. MASC: Massive Arabic Speech Corpus. find
  3. Ali, M.; Elshafei, M.; Al-Ghamdi, M.; Al-Muhtaseb, H.; and Al-Najjar, A. 2008. Generation of arabic phonetic dictionaries for speech recognition. In 2008 International Conference on Innovations in Information Technology, 59–63. find
  4. Baevski, A.; Zhou, H.; Mohamed, A.; and Auli, M. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. arXiv:2006.11477. ↗ our card
  5. Gayraud, F.; Barkat-Defradas, M.; Lahrouchi, M.; and Ben Hamed, M. 2018. Development of phonetic complexity in Arabic, Berber, English and French. Canadian Journal of Linguistics / Revue canadienne de linguistique, 63(4): 527– 555. find
  6. Gibson, M. L. 1999. Dialect contact in Tunisian Arabic: sociolinguistic and structural aspects. Ph.D. thesis, University of Reading. find
  7. Grosman, J. 2021. Fine-tuned XLSR-53 large model for speech recognition in Arabic. https://huggingface.co/ jonatasgrosman/wav2vec2-large-xlsr-53-arabic. ↗
  8. Habash, N.; Soudi, A.; and Buckwalter, T. 2007. On Arabic Transliteration, 15–22. Dordrecht: Springer Netherlands. ISBN 978-1-4020-6046-5. find
  9. Hennequin, R.; Khlif, A.; Voituret, F.; and Moussallam, M. 2020. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50): 2154. Deezer Research. find
  10. Hsu, W.-N.; Bolte, B.; Tsai, Y.-H. H.; Lakhotia, K.; Salakhutdinov, R.; and Mohamed, A. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. arXiv:2106.07447. ↗ our card
  11. Kim, J.; Kong, J.; and Son, J. 2021. Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech. CoRR, abs/2106.06103. ↗
  12. Kong, J.; Kim, J.; and Bae, J. 2020. HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis. arXiv:2010.05646. ↗
  13. Mdhaffar, S.; Bougares, F.; de Mori, R.; Zaiem, S.; Ravanelli, M.; and Estève, Y. 2024. TARIC-SLU: A Tunisian Benchmark Dataset for Spoken Language Understanding. In Calzolari, N.; Kan, M.-Y.; Hoste, V.; Lenci, A.; Sakti, S.; and Xue, N., eds., Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 15606– 15616. Torino, Italia: ELRA and ICCL. find
  14. Ouerhani, B. 2009. Interférence entre le dialectal et le littéral en Tunisie: Le cas de la morphologie verbale. Synergies Tunisie n, 1: 75–84. find
  15. Peddinti, V.; Povey, D.; and Khudanpur, S. 2015. A time delay neural network architecture for efficient modeling of long temporal contexts. In Interspeech. find
  16. Povey, D.; Ghoshal, A.; Boulianne, G.; Burget, L.; Glembek, O.; Goel, N.; Hannemann, M.; Motlı́ček, P.; Qian, Y.; Schwarz, P.; Silovský, J.; Stemmer, G.; and Vesel, K. 2011. The Kaldi speech recognition toolkit. IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. find
  17. Radford, A.; Kim, J. W.; Xu, T.; Brockman, G.; McLeavey, C.; and Sutskever, I. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356. ↗ our card
  18. Saidi, D. 2007. Typology of motion event in Tunisian Arabic. Proceedings of LingO, 196: 23. find
  19. Talafha, B.; Abuammar, A.; and Al-Ayyoub, M. 2021. Atar: Attention-based LSTM for Arabizi transliteration. International Journal of Electrical and Computer Engineering, 11: 2327–2334. find
  20. Tilmatine, M. 1999. Substrat et convergences: le berbère et l’arabe nord-africain. EDNA, Estudios de dialectologı́a norteafricana y andalusı́, 4: 99–119. find
  21. Tjandra, A.; Singhal, N.; Zhang, D.; Kalinli, O.; Mohamed, A.; Le, D.; and Seltzer, M. L. 2022. Massively Multilingual ASR on 70 Languages: Tokenization, Architecture, and Generalization Capabilities. arXiv:2211.05756. van Niekerk, B.; Carbonneau, M.-A.; Zaidi, J.; Baas, M.; Seute, H.; and Kamper, H. 2022. A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE. voicepaw et al. 2024. SoftVC VITS Singing Voice Conversion Fork. Accessed: YYYY-MM-DD. ↗
  22. Zhu, J.; and Zhang, C. 2021. Performing forced alignment with Wav2vec 2.0. The Journal of the Acoustical Society of America, 150: A357–A357. find
  23. Zribi, I.; Ellouze, M.; Belguith, L. H.; and Blache, P. 2017. Morphological disambiguation of Tunisian dialect. Journal of King Saud University - Computer and Information Sciences, 29(2): 147–155. Arabic Natural Language Processing: Models, Systems and Applications. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/naouara-2025-linto-audio-textual-datasets-train/