A newly released Yorùbá text corpus containing 24,905 hand-verified, tone-marked sentences could strengthen the development of speech tools, text-to-speech systems and other artificial intelligence applications for one of Africa’s most widely spoken languages.
A contemporary source for Yoruba language technology
The ÌròyìnSpeech Text Corpus was presented in a research paper first submitted on 4 October 2026 by Kọ́lá Túbọ̀sún, Anuoluwapo Aremu, Tolúlọpẹ́ Ògúnrẹ̀mí, Iroro Orife and David Ifeoluwa Adelani. The researchers are affiliated with YorubaName.com, Stanford University, Niger-Volta LTI and Mila and McGill University.
The resource contains 275,897 word tokens and 15,687 distinct word types. It is the text component behind an existing 42-hour speech corpus recorded by 80 volunteers and described at the LREC-COLING conference in 2024. The new release makes the carefully prepared written material available as a resource in its own right.
According to the researchers, the project was designed to address a persistent weakness in digital Yorùbá resources. Much of the language material commonly used in multilingual datasets comes from religious translation. That can leave systems better prepared for scriptural language than for contemporary news, education, public services and everyday communication.
ÌròyìnSpeech instead brings together modern, secular material in Standard Yorùbá. About 11,000 sentences were adapted from openly licensed news collections, while roughly 12,000 more were created in-house from translated fiction and newly written prompts. The team also added examples containing less common syllabic nasals so that the corpus would better represent sounds that a natural sample might miss.
Why tone marks and Unicode matter
Yorùbá is tonal, so its diacritics and sub-dots are essential parts of written meaning. A missing or incorrect mark can change pronunciation or produce ambiguity. For AI developers, inconsistent encoding can also cause words that look identical to a reader to be processed as different strings.
Every sentence in the corpus was checked manually for tone marks and sub-dots, adjusted for read-aloud clarity and edited towards a neutral register. The researchers also localised some non-Yorùbá names and places to prevent speakers from switching into a different sound system while recording prompts intended for acoustic modelling.
A second technical review exposed another problem. More than 60 per cent of the original lines contained inconsistent Unicode forms, where visually identical letters were stored in different ways. The team corrected those inconsistencies, removed 96 exact duplicates and documented remaining edge cases so that future researchers can understand the data rather than treating it as a flawless black box.
This is significant beyond the corpus itself. Digital systems may display Yorùbá correctly while silently counting, sorting or matching its letters incorrectly. Tools that ignore combining marks can split words, inflate vocabulary counts or fail to recognise equivalent text. The release therefore offers both useful language data and a practical warning for developers building African language products.
Possible uses for speech, education and access
The authors say the corpus can support automatic restoration of missing diacritics, grapheme-to-phoneme conversion, text-to-speech front ends, tone-marking evaluation and research into Yorùbá orthography. Because its sentences were already validated as recording prompts, another team could also use them to record a different dialect, age group, accent or type of voice without beginning months of text curation from scratch.
That creates possibilities for more natural navigation tools, reading support, educational applications, digital archives and voice services. It may also help developers improve the presence of Yorùbá in devices and platforms that still work best in English and other high-resource languages.
The release nevertheless has clear boundaries. It represents Standard Yorùbá used in Nigerian broadcast media and does not cover regional varieties in Nigeria, Benin or Togo. Its register is based on read-aloud news and adapted fiction, not spontaneous conversation. The localisation of names also means the sentences should not be treated as a factual record of the news from which some material was derived.
The text is available through Zenodo in plain text, CSV and JSON Lines formats under a Creative Commons Attribution-NonCommercial 4.0 licence. The paper explains that the non-commercial restriction was necessary because some source material carrying those terms could not be cleanly separated from the in-house sentences. A subset previously contributed to Mozilla Common Voice remains available there under separate terms.
Why this matters for Yoruba cultural preservation
Language preservation now extends beyond books, broadcasting and classroom teaching. It also depends on whether a language can be searched, spoken, heard and processed accurately by digital systems. A well-curated corpus does not solve every problem, but it gives researchers and product builders a stronger foundation than scraped text with missing tone marks and uncertain quality.
For Yoruba speakers, the work is a reminder that cultural continuity and technological participation are increasingly connected. The next test will be whether universities, startups, educators and public institutions turn resources such as ÌròyìnSpeech into useful tools while respecting the language’s complexity and the corpus licence.
Researchers and developers can read the new ÌròyìnSpeech text corpus paper, review the 2024 speech corpus study and consult the linked dataset before planning new Yorùbá language technology projects.
