Elamur Research⌕
☰
home / papers / 2020 / MASRI-HEADSET: A Maltese Corpus for Speech Recognition

MASRI-HEADSET: A Maltese Corpus for Speech Recognition

Summary

Maltese has about 500,000 speakers and a stable, well-studied orthography, yet it had no speech recognition tools, largely because no transcribed speech data existed for training. Working within the University of Malta's MASRI initiative, researchers created MASRI-HEADSET, the first Maltese corpus built specifically for ASR. Twenty-five Maltese-dominant speakers read text prompts in a quiet room. The speakers were balanced by gender and drawn from the six statistical regions of the islands. The prompts came from Korpus Malti v3.0 and were chosen by a greedy procedure that maximised phonetic trigram variety. The released headset recordings total about eight hours. They were cleaned of stuttering and mispronunciations, and their transcripts were normalised. Baseline acoustic models trained in CMU-Sphinx and Kaldi showed that even this small, clean dataset supports usable recognition, with a Kaldi neural network reaching about 10.6% WER. The resource establishes a starting point for Maltese speech technology, and the authors plan to extend it through Common Voice crowdsourcing, data augmentation and handling of Maltese–English code-switching.

Key points

  • MASRI-HEADSET (MHC) contains 8 h 6 min of speech in 3,864 single-speaker utterances, with 11,503 unique word types, distributed as 16 kHz, 16-bit mono audio.

  • There are 25 speakers (13 female, 12 male), aged 18–31 with a mean age of 23.9. They were selected from 61 volunteers for balance in gender, age and region.

  • Prompts were taken from Korpus Malti v3.0. Text was phonetised with a rule-based G2P, and blocks of about 75 words were added greedily to each sample to maximise phone-trigram coverage. Prompts containing proper names or English words were removed.

  • Sessions were recorded with both a headset and a far-field microphone placed about 3 m away, but only the headset audio is released. Transcripts are lowercased and stripped of punctuation, except hyphens and apostrophes.

  • Word error rates were:

    • Sphinx HMMs: 19.40%
    • Sphinx HMMs with LDA/MLLT: 18.40%
    • Kaldi HMMs: 12.54%
    • Kaldi NNs (nnet4d3): 10.56%

    Sentence error rates ranged from 64.0% down to 37.6%. The language model had a perplexity of about 182 and 87 OOV words.

Interesting ideas

  • I1. The prompt set was built by greedily adding the text block that most increases phone-trigram diversity, which extracts the most variety from a fixed recording budget. — For Mova (learn): Pick sentences for LLM-teacher or manual treebank annotation by greedy coverage of rare deprel/POS trigrams and morphological feature combinations, so each labelled sentence adds the most new structure.
  • I2. The authors deliberately removed prompts with proper names and foreign-language words to get a clean monolingual core, and they treat code-switching as a separate, later challenge. — For Mova (tokens): Keep a clean core evaluation set apart from name-heavy or mixed-language text, so tagger and lemmatizer errors can be attributed to the language itself rather than to foreign or unknown tokens.
  • I3. Results are reported as both word-level and sentence-level error, and the much harsher sentence metric exposes how far a system is from producing fully correct outputs. — For Mova (tree): Report whole-sentence exact-match tree accuracy next to LAS, since downstream frames and the world model break whenever any single arc is wrong.
  • I4. Each session was recorded at the same time in clean (close) and degraded (far-field) conditions, but only the clean channel was released first, leaving a parallel harder version for later robustness work. — For Mova (learn): Store paired clean and noisy variants of the same sentences (typos, informal spelling, missing punctuation) to measure how much each pipeline stage degrades, using the clean parse as a reference.
  • I5. For crowdsourced data, the authors plan to rank text–speech pairs automatically by how likely the recording matches its prompt before accepting them. — For Mova (learn): Rank silver parses by witness-parser agreement or model likelihood and admit only the top band, routing low-ranked sentences to the LLM teacher tier instead.

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

14 papers of our library cite it (of the 140884 whose bibliographies we read); 3 of them have a page here, 11 not yet.

Cite

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

@inproceedings{mena2020masri,
  title = {{MASRI-HEADSET: A Maltese Corpus for Speech Recognition}},
  author = {Mena, Carlos Daniel Hernandez and Gatt, Albert and DeMarco, Andrea and Borg, Claudia and Plas, Lonneke van der and Muscat, Amanda and Padovani, Ian},
  booktitle = {Proceedings of the Twelfth Language Resources and Evaluation Conference},
  year = {2020},
  url = {https://aclanthology.org/2020.lrec-1.784/}
}

References

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

  1. Cheng, Q., Chen, G., et al. (2016). Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173–182. find our card
  2. Azzopardi-Alexander, M. and Borg, A. (2013). Maltese. Routledge. find
  3. Besacier, L., Barnard, E., Karpov, A., and Schultz, T. (2014). Automatic speech recognition for under-resourced languages: A survey. Speech Communication, 56:85–100. find
  4. Borg, C. and Gatt, A. (2017). Morphological analysis for the Maltese language: The challenges of a hybrid system. In Proceedings of the 3rd Arabic Natural Language Processing Workshop (WANLP’17), Valencia, Spain. Association for Computational Linguistics. find our card
  5. Borg, M., Bugeja, K., Vella, C., Mangion, G., and Gafà, C. (2011). Preparation of a free-running text corpus for Maltese concatenative speech synthesis. In Proceedings of the 3rd International Conference on Maltese Linguistics (GhiLM’11), Valletta. GhiLM. find
  6. Brincat, J. (2011). Maltese and other languages: A linguistic history of Malta. Midsea Books, Malta. find
  7. Camilleri, J. J. (2013). A Computational Grammar and Lexicon for Maltese. Master’s thesis, Chalmers University of Technology, Gothenburg, Sweden. find
  8. Čéplö, S. (2018). Constituent order in Maltese: A quantitative analysis. Ph.D. thesis, Charles University, Prague, Czech Republic. find
  9. Collobert, R., Puhrsch, C., and Synnaeve, G. (2016). Wav2letter: an end-to-end convnet-based speech recognition system. arXiv preprint arXiv:1609.03193. ↗
  10. Comrie, B. (2009). Maltese and the world atlas of language structures. Introducing Maltese Linguistics, pages 3–12. find
  11. Draxler, C. and Jänsch, K. (2004). Speech recorder (version 2.2. 1). https://www.bas.uni-muenchen.de/Bas/software/speechrecorder/. ↗
  12. European Parliament. (2018). Language equality in the digital age (European Parliament Resolution 2018/2028(INI)). http://www.europarl.europa.eu/doceo/document/TA-8-2018-0332 EN.pdf. ↗
  13. Fabri, R. (1993). Kongruenz und die Grammatik des Mal-tesischen. Niemeyer, Tuebingen. find
  14. Gatt, A. and Čéplö, S. (2013). Digital corpora and other electronic resources for Maltese. find
  15. Corpus Linguistics 2013, page 96. find
  16. Gatt, A. and Fabri, R. (2018). Borrowed affixes and morphological productivity: A case study of two Maltese nominalisations. In P Paggio et al., editors, Languages of Malta: Recent Studies, pages 143–170. Language Science Press, Berlin. find
  17. Grech, S. (2015). Variation in English: Perception and patterns in the identification of Maltese English. Ph.D. thesis, University of Malta, Malta. find
  18. Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., Prenger, R., Satheesh, S., Sengupta, S., Coates, A., and Ng, A. Y. (2014). Deep Speech: Scaling up end-to-end speech recognition. arXiv, 1412.5567. find our card
  19. Hoberman, R. D. (2007). Maltese morphology. Morphologies of Asia and Africa, 1:257–281. find
  20. Huggins-Daines, D., Kumar, M., Chan, A., Black, A. W., Ravishankar, M., and Rudnicky, A. I. (2006). Pocketsphinx: A free, real-time continuous speech recognition system for hand-held devices. In 2006 IEEE International Conference on Acoustics Speech and Signal Processing Proceedings, volume 1, pages I–I. IEEE. find
  21. Krauwer, S. (2003). The basic language resource kit (blark) as the first milestone for the language resources roadmap. Proceedings of SPECOM 2003, pages 8–15. find
  22. Lamere, P., Kwok, P., Gouvea, E., Raj, B., Singh, R., Walker, W., Warmuth, M., and Wolf, P. (2003). The CMU SPHINX-4 speech recognition system. In IEEE Intl. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2003), Hong Kong, volume 1, pages 2–5. find
  23. Lim, M. and Kim, J.-H. (2019). Integration of tensorflow based acoustic model with kaldi wfst decoder. arXiv preprint arXiv:1906.11018. ↗
  24. Matras, Y. (2009). Language Contact. Cambridge University Press, Cambridge, UK. find
  25. Mena, C. D. H., Ruiz, I. V. M., and Camacho, J. A. H. (2017). Automatic speech recognizers for mexican spanish and its open resources. Journal of Applied Research and Technology, 15(3). find
  26. Mena, C. (2019). The ciempiess proper-names pronouncing dictionary. In Corpus presented at OpenCor 2019 Conference, Guanajuato City, Mexico.[Available online at https://opencor.gitlab.io/corpora-list/]. ↗
  27. Micallef, P. (1997). A text to speech synthesis system for Maltese. Ph.D. thesis, University of Surrey. find
  28. Mifsud, M. (1995). Loan verbs in Maltese: A descriptive and comparative study. Brill, Leiden. find
  29. NSO. (2019). Regional Statistics (Malta), 2019 Edition. National Statistics Office, Malta. find
  30. Panayotov, V., Chen, G., Povey, D., and Khudanpur, S. (2015). Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210. IEEE. find
  31. Park, D. S., Chan, W., Zhang, Y., Chiu, C.-C., Zoph, B., Cubuk, E. D., and Le, Q. V. (2019). SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition. arXiv. find our card
  32. Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., Hannemann, M., Motlicek, P., Qian, Y., Schwarz, P., et al. (2011). The Kaldi speech recognition toolkit. In IEEE 2011 workshop on automatic speech recognition and understanding. IEEE Signal Processing Society. find
  33. Ravishankar, V., Tyers, F., and Gatt, A. (2017). A Morphological analyser for Maltese. Procedia Computer Science (Special Issue on Arabic Computational Linguistics), 117:175–182. find
  34. Rosner, M. and Joachimsen, J. (2012). The Maltese Language in the Digital Age. Springer, Berlin and Heidelberg. find
  35. Roter, G. (2019). Sharing our Common Voices – Mozilla releases the largest to-date public domain transcribed voice dataset. https://blog.mozilla.org/blog/2019/02/28/sharing-our-common-voices-mozilla-releases-the-largest-to-date-public-domain-transcribed-voice-dataset/, February. ↗
  36. Schneider, S., Baevski, A., Collobert, R., and Auli, M. (2019). wav2vec: Unsupervised Pre-training for Speech Recognition. In Proceedings of Interspeech 2019, pages 1–9, Graz, Austria. find
  37. Tiedemann, J. and van der Plas, L. (2016). Bootstrapping a Dependency Parser for Maltese - A Real-World Test Case. In Martijn Weiling, et al., editors, From Semantics to Dialectometry : Festschrift in honor of John Nerbonne. College Publications, London. find
  38. Vella, A. (1994). Prosodic structure and intonation in Maltese and its influence on Maltese English. Ph.D. thesis, University of Edinburgh, Scotland. find
  39. Wang, Y., Chen, T., Xu, H., Ding, S., Lv, H., Shao, Y., Peng, N., Xie, L., Watanabe, S., and Khudanpur, S. (2019). Espresso: A fast end-to-end neural speech recognition toolkit. arXiv preprint arXiv:1909.08723. ↗
  40. Weide, R. L. (1998). The CMU pronouncing dictionary. http://svn.code.sf.net/p/cmusphinx/code/trunk/cmudict/. ↗
  41. Young, S. J. and Young, S. (1993). The HTK hidden Markov model toolkit: Design and philosophy. University of Cambridge, Department of Engineering Cambridge, England. find
  42. Zammit, A. (2018). A dependency parser for the Maltese language using deep neural networks. Master’s thesis, Department of Artificial Intelligence, University of Malta, Msida, Malta. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mena-2020-masri-headset-maltese-corpus-speech/