Elamur Research⌕
☰
home / papers / 2024 / TunArTTS: Tunisian Arabic Text-To-Speech Corpus

TunArTTS: Tunisian Arabic Text-To-Speech Corpus

Paper · 2024 · Imen Laouirine, Rami Kammoun and Fethi Bougares

Summary

Tunisian Arabic is a low-resource dialect, and before this work it had no text-to-speech research and no corpus suited to building synthesis systems. The authors searched for a usable source and settled on Derja.Ninja, an online English–Tunisian dictionary released under CC-BY-SA 4.0. Each entry has a recording of a word and a recording of a sentence that uses it.

They scraped about 44 hours of multi-speaker audio and filtered it by sampling rate and other TTS-specific quality requirements. The result is TunArTTS, a single-speaker corpus of just over 3 hours, recorded by a male voice at 44.1 kHz, with manually diacritized transcripts.

On this corpus they trained Tacotron2 and FastSpeech2 baselines in two ways: from scratch, and by fine-tuning models pre-trained on the English LJSpeech dataset. They scored the systems with both listener judgments and objective measures. Transfer learning worked best, reaching a MOS of 3.88. The release of the corpus and a baseline demo makes Tunisian TTS research possible for the first time.

Key points

  • Source and filtering pipeline:
    • A Python script harvested Derja.Ninja, yielding more than 44 hours of multi-speaker audio at four sampling rates (8, 16, 44 and 48 kHz).
    • Keeping only the 44/48 kHz files left about 15.5 hours.
    • This was then reduced to a mono-speaker set of a little over 3 hours.
  • Transcripts: all transcripts were diacritized by hand to reduce ambiguity. Making them CODA-compliant was deferred to future work, because no automatic tooling exists and the manual effort would be large.
  • Data requirements: the paper lists the requirements a TTS corpus should meet, including:
    • a sampling rate of at least 22,050 Hz;
    • clips of 2–10 seconds;
    • silence removed, mono channel, no background noise;
    • consistent tempo and tone;
    • normalized symbols and coverage of all phonemes.
  • Models: Tacotron2 and FastSpeech2, each trained from scratch and with transfer learning from LJSpeech. The best result was a MOS of 3.88, obtained with LJSpeech transfer.
  • Context and release: related Arabic TTS corpora are ASC (about 4 h), ClArTTS (12 h of classical Arabic) and QasrTTS (about 1 h). The release is at github.com/elyadata/TunArTTS.

Interesting ideas

  • I1. Instead of noisy, inconsistent web video, the authors used a dictionary whose entries pair a headword with an example sentence, which gives cleaner, aligned material for a scarce language.
  • I2. They wrote down an explicit checklist of what makes a sample fit for training and filtered a large harvest down to a small, uniform subset that passes it.
  • I3. Fine-tuning a model pre-trained on a large, well-resourced language beat training from scratch on 3 hours of target data.
  • I4. Ambiguity in the written input (missing vowels) was removed by manual annotation before training, so the model never has to guess hidden information.
  • I5. Training strategies were compared using both human opinion scores and automatic metrics, not one alone.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tokens): dictionary example sentences could become lemma-in-context checks for the tagger and lemmatizer, with each headword expected to come out right inside its own example.
  • I2 (learn): silver data could pass a similar written admission checklist (sentence length, witness parser agreement, no unknown tokens), so the reason any sentence was kept or dropped can be explained.
  • I3 (tree): gparse could be pre-trained on large standard English treebanks and then fine-tuned on small sets for informal, dialectal or domain-specific English, rather than trained from zero on those sets.
  • I4 (tokens): a gold layer that explicitly marks lemma/POS for highly ambiguous homographs could serve as unambiguous training input and as a regression set.
  • I5 (learn): parser or reader variants could be judged with both automatic scores and human or LLM-teacher judgments, flagging cases where the two disagree.

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, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksSpeech & spoken language · 480 works

Cite

Cite as Imen Laouirine, Rami Kammoun and Fethi Bougares. 2024. TunArTTS: Tunisian Arabic Text-To-Speech Corpus. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024).

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

@inproceedings{laouirine2024tunartts,
  title = {{TunArTTS: Tunisian Arabic Text-To-Speech Corpus}},
  author = {Imen Laouirine, Rami Kammoun and Fethi Bougares},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year = {2024},
  doi = {10.63317/32ro7rpyrsaw},
  url = {https://doi.org/10.63317/32ro7rpyrsaw}
}

References

42 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Majid Asgari-Bidhendi, Behrouz Minaei, and Hosein Jozi. 2012. Extracting person names from ancient islamic arabic texts. ↗
  2. Massa Baali, Tomoki Hayashi, Hamdy Mubarak, Soumi Maiti, Shinji Watanabe, Wassim El-Hajj, and Ahmed Ali. 2023. Unsupervised data selection for tts: Using arabic broadcast news as a case study. ↗
  3. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. ↗
  4. Evelina Bakhturina, Vitaly Lavrukhin, and Boris Ginsburg. 2022. A toolbox for construction and analysis of speech datasets. ↗
  5. Julia Cambre, Jessica Colnago, Jim Maddock, Janice Tsai, and Jofish Kaye. 2020. Choice of voices: A large-scale evaluation of text-to-speech voice quality for long-form content. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, CHI ’20, page 1–13, New York, NY, USA. Association for Computing Machinery. find
  6. Miloš Cerňak and Milan Rusko. 2005. An evaluation of a synthetic speech using the pesq measure. find
  7. Mingjian Chen, Xu Tan, Yi Ren, Jin Xu, Hao Sun, Sheng Zhao, Tao Qin, and Tie-Yan Liu. 2020. Multispeech: Multi-speaker text to speech with transformer. find
  8. Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. 2015. Attention-based models for speech recognition. Advances in neural information processing systems, 28. ↗
  9. Sara Durrani and Muhammad Arshad. 2021. Transfer learning based speech affect recognition in urdu. ↗
  10. Fady Fahmy, Mahmoud Khalil, and Hazem Abbas. 2020. A transfer learning end-to-end arabictext-to-speech (tts) deep architecture. find
  11. Daniel Griffin and Jae Lim. 1984. Signal estimation from modified short-time fourier transform. IEEE Transactions on acoustics, speech, and signal processing, 32(2):236–243. ↗
  12. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented transformer for speech recognition.
  13. Nizar Habash, Fadhl Al-Eryani, Salam Khalifa, Owen Rambow, Dana Abdulrahim, Alexander Erdmann, Reem Faraj, Wajdi Zaghouani, Houda Bouamor, Nasser Zalmout, Sara Hassan, Faisal Alshargi, Sakhar Alkhereyf, Basma Abdulkareem, Ramy Eskander, Mohammad Salameh, and Hind Saddiki. 2019. Unified guidelines and resources for arabic dialect orthography.
  14. Nawar Halabi. 2016. Modern standard Arabic phonetics for speech synthesis. Ph.D. thesis, University of Southampton. find
  15. Nawar Halabi and Mike Wald. 2016. Phonetic inventory for an arabic speech corpus.
  16. Takaaki Hori, Shinji Watanabe, Yu Zhang, and William Chan. 2017. Advances in joint ctc-attention based end-to-end speech recognition with a deep cnn encoder and rnn-lm. find
  17. Sahar Jamal, Sadaf Abdul Rauf, and Quratulain Majid. 2022. Exploring transfer learning for Urdu speech synthesis. In Proceedings of the Workshop on Resources and Technologies for Indigenous, Endangered and Lesser-resourced Languages in Eurasia within the 13th Language Resources and Evaluation Conference, pages 70–74, Marseille, France. European Language Resources Association. find
  18. Minchan Kim, Myeonghun Jeong, Byoung Jin Choi, Sunghwan Ahn, Joun Yeop Lee, and Nam Soo Kim. 2022. Transfer learning framework for low-resource text-to-speech using a large-scale unlabeled speech corpus. In Interspeech 2022. ISCA. find
  19. Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in Neural Information Processing Systems, 33:17022–17033. find
  20. R. Kubichek. 1993. Mel-cepstral distance measure for objective speech quality assessment. In Proceedings of IEEE Pacific Rim Conference on Communications Computers and Signal Processing, volume 1, pages 125–128 vol.1. find
  21. Ajinkya Kulkarni, Atharva Kulkarni, Sara Abedal-monem Mohammad Shatnawi, and Hanan Aldarmaki. 2023. Clartts: An open-source classical arabic text-to-speech corpus.
  22. Kundan Kumar, Rithesh Kumar, Thibault De Boissiere, Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre De Brebisson, Yoshua Bengio, and Aaron C Courville. 2019. Melgan: Generative adversarial networks for conditional waveform synthesis. Advances in neural information processing systems, 32. ↗
  23. Naihan Li, Shujie Liu, Yanqing Liu, Sheng Zhao, Ming Liu, and Ming Zhou. 2019. Neural speech synthesis with transformer network. find
  24. Abir Masmoudi, Fethi Bougares, Mariem Ellouze, Yannick Estève, and Lamia Belguith. 2018. Automatic speech recognition system for tunisian dialect. Language Resources and Evaluation, 52. ↗
  25. Hala Al Masri and Muhy Eddin Za’ter. 2022. Arabic text-to-speech (tts) data preparation. ↗
  26. Salah Mejri, Mosbah Said, and Inès Sfar. 2009. Pluringuisme et diglossie en tunisie. Synergies Tunisie, 1:53–74. find
  27. Hamdy Mubarak, Amir Hussein, Shammur Absar Chowdhury, and Ahmed Ali. 2021. Qasr: Qcri aljazeera speech resource – a large scale annotated arabic speech corpus. ↗
  28. Jamila Oueslati. 2021. French loans in tunisian arabic from phonetic and phonological perspective. Rocznik Orientalistyczny/Yearbook of Oriental Studies, T. LXXIV(No 1):95–113. ↗
  29. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. Scaling speech technology to 1,000+ languages. ↗
  30. Pascal Puchtler, Johannes Wirth, and René Peinl. 2021. Hui-audio-corpus-german: A high quality tts dataset. find
  31. Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2022. Fastspeech 2: Fast and high-quality end-to-end text to speech. find
  32. Sebastian Ruder, Matthew E. Peters, Swabha Swayamdipta, and Thomas Wolf. 2019. Transfer learning in natural language processing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Tutorials, pages 15– 18, Minneapolis, Minnesota. Association for Computational Linguistics.
  33. Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. 2018. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions. find
  34. Brad Story. 2019. History of speech synthesis, pages 9–33. find
  35. Xu Tan, Jiawei Chen, Haohe Liu, Jian Cong, Chen Zhang, Yanqing Liu, Xi Wang, Yichong Leng, Yuanhao Yi, Lei He, Frank Soong, Tao Qin, Sheng Zhao, and Tie-Yan Liu. 2022. Naturalspeech: End-to-end text to speech synthesis with human-level quality. find
  36. Xu Tan, Tao Qin, Frank Soong, and Tie-Yan Liu. 2021. A survey on neural speech synthesis. arXiv preprint arXiv:2106.15561. ↗
  37. Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, and Rif A. Saurous. 2017. Tacotron: Towards end-to-end speech synthesis. ↗
  38. Jin Xu, Xu Tan, Yi Ren, Tao Qin, Jian Li, Sheng Zhao, and Tie-Yan Liu. 2020. Lrspeech: Extremely low-resource speech synthesis and recognition. ↗
  39. Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. 2020. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP, pages 6199–6203. IEEE. ↗
  40. Inès Zribi, Rahma Boujelbane, Abir Masmoudi, Mariem Ellouze, Lamia Belguith, and Nizar Habash. 2014. A conventional orthography for Tunisian Arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 2355–2361, Reykjavik, Iceland. European Language Resources Association (ELRA). 10.
  41. Language Resource References Honnet, Pierre-Edouard and Lazaridis, Alexandros and Garner, Philip N. and Yamagishi, Junichi. 2017. The SIWIS French Speech Synthesis Database – Design and recording of a high quality French database for speech synthesis. Idiap-RR-03-2017. find
  42. Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/laouirine-2024-tunartts-tunisian-arabic-text-speech/