Ebonyi's languages are not Igbo dialects, and why that changes everything

Ask most Nigerians what language people speak in Ebonyi State, and the answer comes quickly: Igbo. Ask a linguist the same question and they will pause. The Ebonyi cluster sits at the eastern fringe of the Igbo-speaking world, and its constituent languages share some vocabulary with Igbo, but shared vocabulary is not the same thing as belonging to the same language family. Four of the most widely spoken languages in the state have a different story to tell.

What the classifiers actually say

Izii, Ezaa, Ikwo, and Mgbo (sometimes written Ngbo) are classified by Ethnologue and Wikipedia's Izi article as a distinct sub-branch within the Niger-Congo family, not as Igbo dialects. The four form what linguists call the Izii-Ezaa-Ikwo-Mgbo cluster. They share structural features with each other that they do not share with standard Igbo, and their speakers are aware of the distance. A speaker of Izzi arriving in Nnewi or Onitsha can follow some conversation, but extended communication breaks down. Mutual intelligibility with Igbo is marginal, not functional.

The Wikipedia article on Ezaa states the same point plainly: the language is "not mutually intelligible with Igbo." A linguistic and geographical survey of Ebonyi State conducted by researchers documents the same boundary: the cluster is distinct, not transitional. These are not dialects trending toward a standard; they are languages that happen to share a geographic neighbor.

Then there is Korring (also written Orring), spoken in parts of Ebonyi's northern fringe. Korring is classified in a separate sub-branch entirely, not even within the broader Igboid grouping. The Orring Wikipedia article places it in the Cross River branch of Atlantic-Congo. That is not a minor distinction. Korring's nearest linguistic relatives are Cross River languages to the east, not Igbo to the west. It shares the same political state as the Izii-Ezaa-Ikwo-Mgbo cluster, but not their language family. Korring has no documented audio corpus, no NLP tools, and no formal dataset anywhere in the public record.

Why the confusion persists

The misclassification is not malicious, just understandable. Ebonyi State was carved from Enugu and Cross River states in 1996. It sits in the southeast, a region Nigerians and international observers routinely associate with Igbo. Administrative geography becomes linguistic assumption. State-level affiliation collapses into language identity, and the internal diversity of the cluster disappears in the account.

It also persists because the cluster has very little public-facing documentation. Izzi alone has around 540,000 speakers, comparable in size to Norwegian. No trained ASR (automatic speech recognition) model exists for it. There is no speech corpus on Hugging Face, no Common Voice dataset, no published vocabulary. When a language is invisible to technology, the dominant regional language fills the gap in people's minds.

Why it matters for AI

Classification is not just an academic question. When a language is coded as a dialect of a better-resourced neighbor, every engineering decision that follows is wrong.

An Igbo ASR model trained on Onitsha or Owerri speech will perform poorly on Izii and worse on Ezaa. Not because the model is bad, but because the acoustic features, tonal patterns, and vocabulary are different enough that transfer learning gives little benefit. The pattern is documented for other West African clusters. A 2024 paper from the Ibom NLP project (arXiv:2511.06531) on Ibom languages in Akwa Ibom State shows exactly this: languages grouped administratively under a regional umbrella have distinct phonological profiles that require their own data. Treating them as dialects produces systems that fail speakers in practice.

The stakes scale up fast. Nigeria has over 500 languages, of which fewer than 1% appear in any NLP research at all. Every language that gets tagged as a dialect of its neighbor loses its claim to dedicated data collection. It loses its slot in research proposals. Developers do not build tools for dialects; they extend tools built for the parent language. The mislabel is self-reinforcing. The language stays invisible, which confirms to the next researcher that it is just a dialect, which perpetuates the absence.

Fine-tuning a model on native speech changes this dramatically. A recent study on NaijaVoices (arXiv:2505.20564) found that fine-tuning Whisper on locally recorded Nigerian speech cut word error rates by approximately 76%. The model improved not because the architecture changed, but because the data finally matched the speakers. That gap between a generic model and a fine-tuned one is exactly what correct classification and dedicated data collection can close.

The record that already exists

There is a proof point that these languages can carry full literary register. The New Testaments for Izii, Ezaa, and Ikwo were completed in 1980 by Bible translation teams. YouVersion carries the Izii edition, the Ezaa edition, and the Ikwo edition. The Global Recordings Network has free audio scripture for Izii, Ezza, and Ikwo. Decades of careful linguistic and translation work established that these are complete, living languages with their own orthographies and tonal systems. The research community has been slow to catch up.

What we are doing about it

We are building the first open speech corpora for the Ebonyi cluster, starting with Izii and Ezaa. Our approach follows the BOLD (Broad Open Language Data) methodology documented by researchers at Aikuma and refined through community-led collection projects across Africa. We work with speakers directly, with churches that already hold audio archives in these languages, and with traditional institutions that can mobilize their communities for recording days.

We are also engaged with Masakhane, the pan-African NLP research community, and with the Linguistic Association of Nigeria, which runs Language Documentation Institute training. We are in contact with Lanfrica and ELAR for archival deposit. When the first dataset is ready for release, it will be listed on Hugging Face with its correct language codes, its correct classification, and full metadata for every recording.

Changing how a language is classified in a model training pipeline is a small thing. Getting there requires convincing speakers to record, persuading institutions to share archives, building the transcription capacity to turn audio into text, and publishing datasets in formats that developers can actually use. That is the work. We are doing it because no one else is, and because the alternative is a future where Izii and Ezaa and Ikwo and Mgbo and Korring are simply absent from every AI system their speakers will ever encounter.