--- title: "Tigrinya Number Verbalization — Gaim 2026" description: "Tigrinya has about 10 million speakers but few computational resources." authors: ["Fitsum Gaim", "Issayas Tesfamariam"] year: 2026 venue: "arXiv:2601.03403" source: "https://arxiv.org/abs/2601.03403" doi: "10.48550/arxiv.2601.03403" arxiv: "2601.03403" licence: "CC BY-SA 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by-sa/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/gaim-2026-tigrinya-number-verbalization-rules-algorithm/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gaim-2026-tigrinya-number-verbalization-rules-algorithm/) [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/) / Tigrinya Number Verbalization: Rules, Algorithm, and Implementation # Tigrinya Number Verbalization: Rules, Algorithm, and Implementation **Authors:** [Fitsum Gaim](https://md.research.elamur.ai/free/en/authors/fitsum-gaim/), Issayas Tesfamariam · **Year:** 2026 · **Venue:** arXiv:2601.03403 · **Licence:** [CC BY-SA 4.0 (arXiv)](http://creativecommons.org/licenses/by-sa/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2601.03403) · [PDF](https://arxiv.org/pdf/2601.03403) · [DOI 10.48550/arxiv.2601.03403](https://doi.org/10.48550/arxiv.2601.03403) · [arXiv 2601.03403](https://arxiv.org/abs/2601.03403) · [OpenAlex](https://openalex.org/W7119497821) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) ## Summary Tigrinya has about 10 million speakers but few computational resources. Common tools such as num2words do not support it, so systems like TTS and ASR have no reliable way to turn digits into spoken words. The authors write down the standard rules for saying numbers aloud: cardinals, ordinals, decimals, negatives, percentages, dates, times, currency and phone numbers. The hardest parts are a linking suffix ን (/n/, "and"), the exceptions for teens and hundreds, and when scale words take that suffix. They turn the cardinal rules into an explicit algorithm and release it as an open-source Python package, Tigrinya-Numbers, with unit tests. They also built a 100-item test set focused on difficult cases and used it to check six frontier LLMs. The best model, Opus 4.5, got only 65% exactly right, and the others did far worse. The authors conclude that written-down rules and deterministic code are still needed for production pipelines in low-resource languages. ## Key points - **The linking suffix ን (/n/):** - Single digits never take it. - In compound numbers, every component takes it. - Teens (11–19) take no internal suffix, but do take the final one inside a larger number. - **Hundred changes form:** standalone ሚእቲ (/mi'ti/) becomes ሚእትን (/mi'tn/) inside compounds. Scale words such as thousand or million also take the suffix, unless they cover the whole number at their level. - **Ordinals:** 1st–10th are separate, gendered words. From 11th on, the form is መበል (/mäbäl/) plus the cardinal. - **Package and algorithm:** Tigrinya-Numbers (github.com/fgaim/tigrinya-numbers) covers seven categories and has flags for alternative forms. Examples are ዜሮ vs ባዶ for zero, and reading phone numbers in pairs or single digits. Its tests cover cardinals up to 10^24, and it defaults to the Eritrean spelling variant. - **LLM results:** scoring was exact match after Unicode normalization. | Model | Overall accuracy | |---|---| | Opus 4.5 | 65% | | Gemini 3 Flash | 44% | | Gemini 3 Pro | 31% | | Sonnet 4.5 | 20% | | GPT-5.2 | 17% | | GPT-5 Mini | 1% (mostly ran out of tokens) | Typical errors were leftover digits or Amharic words, a missing suffix, wrong teen forms, and confusing simple with compound multipliers. ## Interesting ideas - The algorithm splits a number into "parts" that each get the linking morpheme. A single predicate (is the multiplier a teen, a round ten or a round hundred?) decides whether a scale word fuses with its multiplier or stands as its own part, which keeps the rules small and checkable. — For Mova (back): number realization in English generation could use the same parts-plus-joining-rule design, with "and", hyphens and scale words added by one inspectable rule instead of ad-hoc string building. - Some words change shape depending on context: hundred is ሚእቲ alone but ሚእትን in a compound. The algorithm handles this as one final rewrite step that applies only when the result has a single part, instead of spreading special cases through the code. — For Mova (back): context-dependent surface forms (a/an, contractions) could also be applied as a last, logged pass over the generated sequence, so each change stays explainable. - The test set was deliberately loaded with known trouble cases and varied digit choices, so that memorized values cannot pass for real competence. — For Mova (learn): evaluation sets for number and quantity handling could be built the same way, with adversarial value choices acting as a negative control against shallow pattern matching. - The authors grouped LLM failures into named error types: leftover digits, missing conjunction, teen errors and scale-word mistakes. This tells developers which rule is breaking, not just how often outputs fail. — For Mova (learn): comparing LLM-teacher outputs against a deterministic rule oracle and sorting the mismatches into named categories could drive targeted repair rules and decide when a teacher tier can be trusted. - Times, currency and phone numbers are treated as structured values with fixed slots: hour, minute, second; main unit and subunit; digit pairs. The reading is produced from that structure rather than from the raw string. — For Mova (world): numeric expressions in a text could be stored as typed values (time, money with subunit, identifier) in the world model, so later steps reason over amounts instead of token strings. *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{gaim2026tigrinya, title = {{Tigrinya Number Verbalization: Rules, Algorithm, and Implementation}}, author = {Gaim, Fitsum and Tesfamariam, Issayas}, year = {2026}, doi = {10.48550/arxiv.2601.03403}, eprint = {2601.03403}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2601.03403} } ``` - [← previous TigerCoder: A Novel Suite of LLMs for Code Generation in Bangla](https://md.research.elamur.ai/free/en/papers/raihan-2026-tigercoder-novel-suite-llms-code/) - [next → To Overfit or Not to Overfit? An Evaluation of HTR Workflow on 17Th-18Th Century French Corpus](https://md.research.elamur.ai/free/en/papers/tiger-2026-overfit-or-not-overfit-evaluation/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/) ## References 7 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Eberhard, D. M., Simons, G. F., and Fennig, C. D. Ethnologue: Languages of the world, 2024. URL [https://www.ethnologue.com](https://www.ethnologue.com). [↗](https://www.ethnologue.com) 2. Gaim, F. and Park, J. C. Natural Language Processing for Tigrinya: Current State and Future Directions. arXiv preprint, July 2025. doi: 10.48550/arXiv.2507.17974. URL https://arxiv.org/abs/2507.17974. [↗](https://doi.org/10.48550/arxiv.2507.17974) [our card](https://md.research.elamur.ai/free/en/papers/gaim-2025-natural-language-processing-tigrinya-current/) 3. Keletay, M. A. and Worku, H. S. Developing concatenative based text to speech synthesizer for Tigrigna language. Internet of Things and Cloud Computing, 8(2):24–30, 2020. doi: 10.11648/j.iotcc.20200802.12. URL https://doi.org/ 10.11648/j.iotcc.20200802.12. [↗](https://doi.org/10.11648/j.iotcc.20200802.12) 4. Mihreteab, M., Kidane, S., Fisshaye, M., and Teklemariam, F. Towards End-to-End Speech Synthesis for Tigrinya Language. In Chakravarthi, B. R., B, B., Rajiakodi, S., García Cumbreras, M. Á., Jiménez Zafra, S. M., Kovács, G., Eger, S., Wahyu Pamungkas, E., and Dobrovoljc, K. (eds.), Speech and Language Technologies for Low-Resource Languages, pp. 21–35, Cham, Nov 2025. Springer Nature Switzerland. ISBN 978-3-032-05855-3. URL https://link.springer.com/chapter/10.1007/978-3-032-05855-3_2. [↗](https://doi.org/10.1007/978-3-032-05855-3_2) 5. Pratap, V., Tjandra, A., Shi, B., Tomasello, P., Babu, A., Kundu, S., Elkahky, A., Ni, Z., Vyas, A., Fazel-Zarandi, M., Baevski, A., Adi, Y., Zhang, X., Hsu, W.-N., Conneau, A., and Auli, M. Scaling Speech Technology to 1,000+ Languages: Massively Multilingual Speech (MMS) - Tigrigna Text-to-Speech, 2023. URL [https://huggingface.co/facebook/mms-tts-tir](https://huggingface.co/facebook/mms-tts-tir). Part of the Massively Multilingual Speech project. [↗](https://huggingface.co/facebook/mms-tts-tir) [our card](https://md.research.elamur.ai/free/en/papers/pratap-2023-scaling-speech-technology-1-000/) 6. Savoir-faire Linux. num2words: Modules to convert numbers to words, 2024. savoirfairelinux/num2words. [find](https://scholar.google.com/scholar?q=Savoir%2Dfaire+Linux%2E+num2words%3A+Modules+to+convert+numbers+to+words%2C+2024%2E+savoirfairelinux%2Fnum2words%2E) 7. URL [https://github.com/Tesfamariam](https://github.com/Tesfamariam), I. Let’s Speak Tigrinya: A Multidimensional Approach to the Teaching and Learning of Tigrinya as a Foreign Language. Let’s Speak African Language Series. National African Language Resource Center (NALRC) Press, Bloomington, IN, 2018. ISBN 978-1-59703-030-4. A First-Year Textbook. [↗](https://github.com/Tesfamariam)