Elamur Research⌕
☰
home / papers / 2020 / Large Vocabulary Read Speech Corpora for Four Ethiopian Languages: Amharic, Tigrigna, Oromo and Wolaytta

Large Vocabulary Read Speech Corpora for Four Ethiopian Languages: Amharic, Tigrigna, Oromo and Wolaytta

Summary

Speech recognition requires transcribed audio, and most Ethiopian languages have very little of it. The authors at Addis Ababa University built read-speech corpora for four languages from three families: Amharic and Tigrigna (Semitic), Oromo (Cushitic) and Wolaytta (Omotic). Prompt sentences came from web and earlier text collections, which were cleaned and normalized. For the two Ethiopic-script languages, the prompts were chosen with a syllable-based phonetic-balance algorithm. For the two Latin-script languages, a simple length limit was used instead. About 100 speakers per language read individually assigned prompts on Android phones, at times and places they chose themselves. This produced roughly 100 hours of training speech in total. HMM-GMM baseline recognizers reached word error rates between about 31% and 38%, so the data is usable for further research. The analysis shows that vocabulary coverage (OOV rate), which depends on how complex each language's morphology is, mattered more for accuracy than language-model perplexity. The procedure is meant as a template for the 70+ other Ethiopian languages that still lack such resources.

Key points

  • Training speech per language: Amharic 26.5 h, Tigrigna 22.1 h, Oromo 22.8 h, Wolaytta 29.7 h.
  • Baseline WERs: Amharic 37.65%, Tigrigna 31.03%, Oromo 38.02%, Wolaytta 33.89%.
  • Recording setup:
    • Six TECNO K7 phones running the Lig-Aikuma recording app.
    • Mono audio, 16 kHz, 16-bit PCM.
    • Semi-supervised sessions in which speakers picked their own time and place.
  • Prompt allocation:
    • Each of 100 speakers got 130 distinct, randomly drawn prompts for Amharic, Tigrigna and Oromo, aiming for at least 100 clean utterances each.
    • Wolaytta speakers got 150 prompts, because speakers were harder to recruit.
  • Language profile:
    • The four languages share roughly 70% of their phone inventories, including ejectives.
    • Amharic's phone set is a subset of Tigrigna's.
    • Oromo and Wolaytta are tonal and distinguish long and short vowels.

Interesting ideas

  • I1. The authors explain each language's WER by separating LM perplexity from decoding-vocabulary OOV rate, showing that low perplexity cannot rescue a system with poor vocabulary coverage. — For Mova (learn): Evaluation reports could split errors into "unknown item" versus "known item, wrong decision," so that fixes target either the lexicon or the model.
  • I2. Each speaker was assigned more prompts than the target number of clean recordings, with extra slack where recruitment was hard, to absorb expected losses. — For Mova (learn): When generating silver data with several witness parsers, request more candidates than needed, because disagreement filtering will discard a predictable share.
  • I3. When the phonetic-balance tool did not fit a script, the authors fell back to a cheap, explainable rule (sentences under 20 words) rather than dropping selection entirely. — For Mova (answer): Test-item selection could use principled coverage criteria where available and simple documented filters elsewhere, keeping test suites reproducible.
  • I4. They give every speaker a distinct random prompt set instead of shared sentences, which widens lexical coverage of the corpus. — For Mova (learn): Disjoint sampling of evaluation or self-play items across runs reduces overlap and gives a broader picture of where the parser fails.
  • I5. The authors treat OOV rate and vocabulary growth as a proxy for morphological complexity, tying a linguistic property to measured system behavior. — For Mova (tokens): Tracking OOV rates for surface forms versus lemmas would show where the morphology module actually buys coverage.

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

6 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 5 not yet.

Cite

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

@inproceedings{abate2020large,
  title = {{Large Vocabulary Read Speech Corpora for Four Ethiopian Languages: Amharic, Tigrigna, Oromo and Wolaytta}},
  author = {Abate, Solomon Teferra and Tachbelie, Martha Yifiru and Melese, Michael and Abera, Hafte and Abebe, Tewodros and Mulugeta, Wondwossen and Assabie, Yaregal and Meshesha, Million and Afnafu, Solomon and Seyoum, Binyam Ephrem},
  booktitle = {Proceedings of the Twelfth Language Resources and Evaluation Conference},
  year = {2020},
  doi = {10.18653/v1/2020.winlp-1.5},
  url = {https://doi.org/10.18653/v1/2020.winlp-1.5}
}

References

14 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abate, S. T., Menzel, W., and Tafila, B. (2005). An amharic speech corpus for large vocabulary continuous speech recognition. In INTERSPEECH 2005 - Eurospeech, 9th European Conference on Speech Communication and Technology, Lisbon, Portugal, September 4-8, 2005, pages 1601–1604. find
  2. Abate, S. T., Melese, M., Tachbelie, M. Y., Meshesha, M., Atinafu, S., Mulugeta, W., Assabie, Y., Abera, H., Ephrem, B., Abebe, T., Tsegaye, W., Lemma, A., Andargie, T., and Shifaw, S. (2018). Parallel corpora for bilingual English-Ethiopian languages statistical machine translation. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3102– 3111, Santa Fe, New Mexico, USA, August. Association for Computational Linguistics. find
  3. Abera, H. and Hailemariam, S. (2018). Design of a Tigrinya language speech corpus for speech recognition. In Proceedings of the First Workshop on Linguistic Resources for Natural Language Processing, pages 78–82, Santa Fe, New Mexico, USA, August. Association for Computational Linguistics. find
  4. Abera, H., Nadeu, C., and Mariam, S. H. (2016). Extraction of syllabically rich and balanced sentences for tigrigna language. In 2016 International Conference on Advances in Computing, Communications and Informatics, ICACCI 2016, Jaipur, India, September 21-24, 2016, pages 2094–2097. find
  5. 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
  6. Chen, S. F. and Goodman, J. (1996). An empirical study of smoothing techniques for language modeling. In 34th Annual Meeting of the Association for Computational Linguistics, pages 310–318, Santa Cruz, California, USA. Association for Computational Linguistics. find our card
  7. Gauthier, E., Blachon, D., Besacier, L., Kouarata, G., Adda-Decker, M., Rialland, A., Adda, G., and Bachman, G. (2016). Lig-aikuma: A mobile app to collect parallel speech for under-resourced language studies. In Interspeech 2016, 17th Annual Conference of the International Speech Communication Association, San Francisco, CA, USA, September 8-12, 2016, pages 381–382. find
  8. Griefenow-Mewis, C. (2001). A Grammatical Sketch of Written Oromo. find
  9. Leslau, W. (2000). Introductory grammar of Amharic. Porta linguarum orientalium, neue Serie, Bd. 21. Harrassowitz, Wiesbaden. find
  10. Pellegrini, T. and Lamel, L. (2009). Automatic word decompounding for ASR in a morphologically rich language: Application to amharic. IEEE Trans. Audio, Speech & Language Processing, 17(5):863–873. find
  11. Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., Hannemann, M., Motlicek, P., Qian, Y., Schwarz, P., Silovsky, J., Stemmer, G., and Vesely, K. (2011). The kaldi speech recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society, December. IEEE Catalog No.: CFP11SRW-USB. find
  12. Stolcke, A. (2002). Srilm – an extensible language modeling toolkit. In IN PROCEEDINGS OF THE 7TH INTER-NATIONAL CONFERENCE ON SPOKEN LANGUAGE PROCESSING (ICSLP 2002, pages 901–904. find
  13. Tachbelie, M. Y. and Abate, S. T. (2015). Effect of language resources on automatic speech recognition for amharic. In AFRICON 2015, Addis Ababa, Ethiopia, September 14-17, 2015, pages 1–5. find
  14. Tesfay, T. Y. (2002). A modern grammar of Tigrinya. Tipografia U. Detti, Rome. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abate-2020-large-vocabulary-read-speech-corpora/