JupiteX Get the app
Science & Technology18 Nov 2026 · about 6 min

From tonal markings to XML files: How Ubini is teaching machines to speak Edo language

The brief

Ubini is creating a digital corpus from recordings of Edo speakers. A corpus is a collection of language material gathered in a form that can be preserved and used. The company’s aim is to save cultural knowledge that might otherwise disappear and provide material for language technology and creative work. Its recordings include oral histories, recipes and other cultural heritage. Ubini has interviewed about 20 people and transcribed more than 20 hours of recordings. It has also made Mia Mia, a 27-minute documentary featuring 85-year-old Maria Evbagharu Osazee reflecting on her life and cultural loss. The archive matters because some Edo knowledge has never been written down. Ohumu says her grandmother’s death left family recipes and words she could no longer recover. Ubini hopes its collected material can support AI tools as well as films, books and other creative works, while preserving knowledge for future generations.

01

What is Ubini creating from Edo speakers’ recordings, and what kinds of knowledge will it preserve?

Ubini is creating a digital corpus from recordings of Edo speakers. A corpus is a collection of language material gathered in a form that can be preserved and used. The company’s aim is to save cultural knowledge that might otherwise disappear and provide material for language technology and creative work.

Its recordings include oral histories, recipes and other cultural heritage. Ubini has interviewed about 20 people and transcribed more than 20 hours of recordings. It has also made Mia Mia, a 27-minute documentary featuring 85-year-old Maria Evbagharu Osazee reflecting on her life and cultural loss.

The archive matters because some Edo knowledge has never been written down. Ohumu says her grandmother’s death left family recipes and words she could no longer recover. Ubini hopes its collected material can support AI tools as well as films, books and other creative works, while preserving knowledge for future generations.

02

How much Edo material has Ubini collected so far, and how many contributors does it hope to record?

Ubini’s reported work so far includes about 20 interviews and more than 20 hours of transcribed recordings. These are recordings of Edo speakers, gathered for a digital corpus of oral histories, recipes and other cultural heritage. The material is intended to preserve knowledge and provide useful language data.

The article also describes a 27-minute documentary, Mia Mia. It features Maria Evbagharu Osazee, an 85-year-old woman reflecting on her life and the cultural loss she has witnessed. The company focuses on people over 70 who speak Edo as their first or main language.

The supplied text does not give a target number of contributors Ubini hopes to record. It ends while describing the company’s focus on older speakers and the kinds of recordings it captures. So the figures that can be confirmed are its current interview and transcription totals, not a future recording target.

03

Why can AI models produce convincing-sounding Edo that is actually inaccurate?

AI models need language material to learn how words and phrases are used. The article describes Edo as poorly represented in digital resources, leaving developers with little reliable data for building language tools. A 2026 study identified 41 African languages supported by the models it examined and 23 public datasets; more than 98% of African languages were unsupported.

That shortage can leave models imitating patterns without reliably knowing Edo. Ohumu says existing resources can also be influenced by English. In one test, a model translated “download” as “D-A-N-L-O-D,” reproducing an English pronunciation rather than giving an Edo equivalent. The result could sound plausible to someone who does not know the language.

This makes reliable, verified data important. Edo is tonal, and differences in tone markings can change meaning. If training material is incomplete or inaccurate, models risk repeating errors in translations and generated speech. Ubini is collecting recordings and cultural material to help address that gap.

04

What happens to a word’s meaning in Edo when its tone changes?

Edo is a tonal language, meaning tone is part of how words are distinguished. As the article explains, differences in tone markings can change a word’s meaning. A change in tone is therefore not just a change in how a word sounds; it can affect what the word communicates.

The article does not give a specific Edo word pair to demonstrate this. It does point out that written tone markings matter when preparing language data. If those markings are missing or inaccurate, a computer may not have the information needed to distinguish words correctly.

That creates a challenge for AI systems trained on incomplete material. They may reproduce errors in translations or generated speech. Ohumu argues that reliable Edo data is needed to support more accurate language technology. The same care matters when recording and transcribing the language, so that the information computers receive reflects the distinctions Edo speakers use.

05

What could be lost when a language disappears, and what might reliable Edo data make possible?

Language loss can mean losing more than words. Ohumu says language carries knowledge, and recalled a study about indigenous communities where understanding of a plant’s medicinal value existed in only one language. If that language disappears, the community’s understanding of what the plant can do may disappear with it.

The same concern applies to Edo knowledge that has not been written down. Ohumu’s grandmother took family recipes and words she translated with her when she died. Ubini is collecting oral histories, recipes and other cultural heritage so that knowledge can be preserved and used in the future.

Reliable Edo data could also make technology more useful. Ohumu points to voice-based services for people more comfortable speaking Edo, including financial services that would let market women carry out banking transactions through audio commands. Such applications depend on verified language data, the gap Ubini is working to address.

06

What is a digital language corpus, and how can it help computers learn to understand a language?

A digital language corpus is an organized collection of language material stored in digital form. It may include written text, speech recordings, transcriptions and information describing the material. The supplied article does not define the term, but it describes Ubini’s project as a digital corpus of Edo oral histories, recipes and other cultural heritage.

A computer can use a well-organized collection as examples of how a language is spoken or written. Recordings and transcripts can provide language material for developing or evaluating language tools. Details such as who is speaking or what a recording contains can also help people find and study relevant examples.

Ubini has interviewed about 20 people and transcribed more than 20 hours of recordings. The need is clear: Ohumu says Edo has limited reliable digital material, and AI responses can sound convincing while being wrong. A larger, verified corpus could give developers better source material, though collecting data alone does not guarantee accurate AI.

07

What is XML, and how does structured text help computers store and process language data?

XML, or Extensible Markup Language, is a format for marking information with tags. Those tags identify parts of a document and can nest inside one another, creating a clear structure. The article does not discuss XML; this definition is based on established technical knowledge.

For language data, a structured record could distinguish a transcript from its translation, speaker details or recording information. A computer program can then locate and process each part rather than treating the entire record as an undivided block of text. XML is one possible format for representing such information; it is not the only one.

Structure can make a corpus easier to organize, search and transfer between tools. It does not make a translation accurate or teach a model Edo by itself. The quality still depends on the language material and its labels. That distinction matters for Ubini’s goal: reliable recordings and verified data are essential, whatever format is used to store them.

This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.

Read more in the JupiteX app

Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.

Or read more news on the web