Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

Igbo speech prototype hears tone

Written November 12, 2026

In Igbo, tone changes what a word means, and most speech tools were not built for that. A working prototype at igbo.ai lets you hear one word in its different tone forms, recorded by native speakers, and pick the meaning you intended. It is a small demonstration, and it points at the exact place larger systems fall short.

Hear the tone, choose the meaning

The Igbo Speech Project prototype is built around the project's own Igbo recordings. On the Speech page you can compare pronunciation and meanings, try speech input in the browser, and explore how a simple question can resolve an ambiguous word. The example the site leads with asks which meaning you intended, and presents one word in four tone forms: àkwà, àkwá, ákwà and ákwá. You hear the word, compare the readings, and choose a meaning.

A dictionary lookup through IgboAPI extends the Speech page beyond the project's own small reference collection. The site asks users to keep external lookup results attributed to their sources, and to check the project's recordings, meanings, annotations and permissions as part of reviewing the foundations.

Why tone is the hard part

The project's starting point is that speaking Igbo naturally takes more than reading letters. Tone changes meaning. Diacritics matter. Dialects carry identity. The site says these details too often get lost in today's speech systems, and puts native speakers at the heart of how quality is defined.

Its sketch of the path from writing to voice runs Text, then Tone, then Phonemes, then Voice. The research hypothesis behind the work is a question, stated as one: can explicit tone, phoneme and dialect supervision make Igbo AI speech measurably more native-sounding? The prototype exists to start answering that, not to declare it answered.

A prototype, stated plainly

The site is direct about the stage of the work. In its own words, this is an application prototype, not a trained Igbo AI speech model. What exists today is recorded audio and experimental browser recognition. Generated speech, dialect expansion and model training remain research milestones.

The prototype sits in the project's Experiment Workshop, which now hosts the Igbo Speech Project, with future conversation and evaluation previews alongside it. It is the input side of a problem other tools approach from the output side: ZONOS3-Mini generates speech across many languages, while this prototype works on hearing and sorting out one tonal language.

Behind the prototype is a wider effort: an open language technology initiative led by Galaxyway AI, built in collaboration with linguists, researchers, native speakers, technologists and the global Igbo community. The stated aim is an open foundation for AI that can understand, pronounce and speak Igbo naturally, organized as six building blocks within one long-term research programme.

Evidence before scale

The roadmap the site describes is Phase 0 followed by six evidence-gated phases. The working approach is to use the prototype to learn what people need, review the recordings, and measure the next speech experiment before expanding scope, improving one measured weakness at a time. Two tests are named: whether clarification helps people reach the intended meaning, and how recognition performs on fresh material under native-speaker review.

The next evidence gate is specific: reviewed lexical records, a small independent challenge set, and a reproducible report of clarification and speech-recognition results. Until that gate is passed, the honest summary is the one the site itself gives: a working start.

The project also places its work on top of what already exists. It names Nkọwa okwu, IgboAPI, Masakhane, Common Voice, Google FLEURS, African Voices, Meta Omnilingual ASR and OmniVoice as datasets, models and communities that inform the research landscape, and states plainly that these are not announced partnerships.

Limits

Related playbooks

Sources

Back to: Everyday playbooks hub · Also: ZONOS3-Mini: open, multilingual, real-time voice