Turaco About RC 9251155

Turaco

A text-to-speech system for Nigerian languages, starting with Hausa.

Why voice

Millions of Nigerians are locked out of digital education, health information, and commerce because those systems only speak English. Voice removes the literacy barrier and the language barrier at once.

Where it stands

Language
Hausa
Corpus
WAXAL — Google’s open speech corpus
Licence
CC-BY-4.0
Duration
~13 h single-speaker studio audio [provisional]
Speech type
Read — phonetically balanced scripts, studio, voice actors
Stage
Baseline in training
Next
Conversational data collection

Hausa is first. We’re training on WAXAL, Google’s open speech corpus released under CC-BY-4.0 — roughly 13 hours of single-speaker studio audio for Hausa, comparable in scale to the datasets behind established English TTS systems. Enough for a baseline. Not enough for a conversation.

What exists in the open data is read speech: phonetically balanced scripts, recorded in studios, by voice actors reading. What it lacks is how people actually speak — the hesitations, the register shifts, the code-switching between Hausa and English mid-sentence, the difference between explaining something to a child and explaining it to a stranger.

A system trained only on read speech reads. It doesn’t talk.

What we’re solving next

Conversational nuance.

Collecting and processing high-quality Hausa audio that captures everyday speech rather than scripted speech. This is the work that decides whether Turaco is usable or merely demonstrable.

Turn-taking.

A system that speaks but cannot stop speaking is not a conversation. Detecting when a person begins talking, yielding, and resuming coherently is a separate problem from synthesis, and we’re treating it as one.

Pidgin.

Nigerian Pidgin is absent from every major open speech corpus, including WAXAL. It is also among the most widely spoken languages in the country. That gap will have to be closed by collection, not by fine-tuning.

Then the rest.

Hausa first, perfected, with the process documented. Yoruba, Igbo, and Pidgin follow — one at a time, each learning from the last.

The family

Turaco is the family. Species denote tier.

  • Turaco Violet — open, Apache 2.0. Named for the Violet Turaco, a West African species common in Nigeria. Unrestricted: no user caps, no commercial restrictions, no mandatory branding.
  • Turaco Great Blue — frontier tier. More expressive voices, conversational turn-taking, lower latency, custom voices.

Models are versioned by language coverage, never renamed:

SpectreLabsAI/turaco-violet-tts-ha

What’s open, what isn’t

Terms

Models are open. Inference is available. The conversational data we collect stays with us — it is the work we are doing, and it is what funds the next language. What we publish is what we learn: research, findings, methods, and the openings we can’t take ourselves.

Models
open
Inference
available
Conversational data
stays with us
Published
research, findings, methods

Open

Contact

This work needs native speakers, recording capacity, compute, and people who know these languages better than any dataset does. If that’s you, or if you’re working on the same problem from another direction, we’d like to hear from you.