Asante Twi TTS training now
A VITS voice trained on 54 hours of studio-quality Asante Twi — with a text pipeline we rebuilt verse-by-verse to restore the ɛ and ɔ that upstream datasets had stripped out.
Accra · Kumasi · the cocoa belt
Voice AI that can finally say Onyankopɔn.
Most speech technology silently deletes the letters ɛ and ɔ — and with them, the meaning of what millions of Ghanaians say every day. Neuralis builds text-to-speech, speech recognition and language tools for Twi, Ewe, Ga and Dagbani, then puts them to work in apps for news, farming, learning and health.
ɛ and ɔ are not accents or decorations — they are letters. ɔba (child) is not oba. Our models are trained on text that keeps every letter.
The gapVoice is how development reaches everyone
<2%
of Africa's ~2,000 languages have any meaningful support in today's AI models.
~46%
of mobile users in Sub-Saharan Africa are still on feature phones — voice calls and radio, not apps, are how information travels.
0
Ghanaian languages are supported by Google Cloud, Azure or ElevenLabs speech services today. We think that should change — so we're building it ourselves.
ModelsSpeech models we train ourselves
Owning the models means no vendor can switch our languages off — and every app we ship makes the models better.
Akuapem Twi & Ewe TTS data ready
The same pipeline, applied to Akuapem Twi and Ewe. Datasets are cleaned and staged; training starts when the first run completes.
Twi speech recognition in development
Fine-tuned recognition for real Ghanaian speech — market noise, code-switching, all dialects. Built on a 335-hour Twi/Fante corpus and measured honestly on field recordings, not leaderboard samples.
PlatformOne voice platform, many apps
Every app needs the same plumbing: channels people actually use, speech in and out, translation, safety rails, and honest evaluation. We build it once.
- Channels first. WhatsApp voice notes, ordinary phone calls (IVR) and USSD — because reach matters more than app-store polish.
- Language layer. Text normalisation that speaks numbers and cedi amounts properly in Twi, translation routing, and code-switch handling.
- Safety rails. In health and finance, answers come from expert-reviewed content blocks — the AI arranges them, it never improvises advice.
- Evaluation gates. Nothing ships on vendor claims. Every model passes our own word-error and translation tests before users hear it.
- Consent-first data. Ghana DPA-compliant consent in the user's language, with real deletion on request — and every consented conversation improves the models.
AppsPublic-interest apps, shipped in waves
CivicRead MVP built
News & accessibility · Twi
Send any text on WhatsApp, hear it read aloud in Twi. A daily audio news digest for people who are blind, low-vision, or read better with their ears.
AgriVoice next up
Agriculture · Twi & Fante
Cocoa farmers ask questions out loud — voice note in, spoken answer out — grounded in agronomist-reviewed guidance, with a human extension officer behind every uncertain answer.
ExamPrep planned
Education · Twi voice
BECE & WASSCE tutoring on WhatsApp with explanations spoken in Twi — building on approaches proven to lift scores by 11 percentage points at ~$5 per student.
MoneyVoice planned
Financial inclusion · Twi & Ga
Speak instead of tapping USSD codes. Voice-driven mobile-money assistance for people locked out of digital finance by literacy or sight.
MamaVoice planned · with clinical partners
Maternal health · Twi, Ewe, Dagbani
Spoken antenatal guidance and danger-sign triage with nurse escalation — voice-first, template-only content, built only alongside health professionals.
TeacherKit planned
Education · Twi, Ewe, Dagbani
Turn English lesson plans into local-language audio teachers can use in class the same day.
LanguagesWhere we start, and why
LanguageSpeakersOur status
Twi (Asante & Akuapem)9M+TTS training · recognition in development
Ewe5M+dataset staged · translation source verified
Ga2M+planned — needs new open voice data, which we aim to help create
Dagbani3M+planned — northern-Ghana apps depend on it
Speaker counts are approximate. We publish what works and what doesn't — including the defects we find in public datasets, so the whole field benefits.
Building in the openRecent progress
- Jul 2026First VITS Asante Twi voice is training on our own hardware — 20,815 studio clips, 54 hours.
- Jul 2026Found and fixed a defect in a widely-used open dataset: its Twi transcripts had every ɛ and ɔ silently removed. We rebuilt all 21,000+ transcripts from the original Bible text, verse by verse, and validated 98% of them.
- Jul 2026CivicRead MVP complete: WhatsApp text-to-Twi-audio with consent-first onboarding and real data deletion on request.
- Jul 2026Published an honest baseline of an existing Twi speech-recognition model: not yet usable in the field. That's the point of measuring.
Work with usThis takes partners
We're looking for pilot communities, farmer networks, schools, health organisations, broadcasters with content to localise, and funders who back African-language AI. If that's you, we'd like to talk.
hello@voiceailabs.devDevelopers: our language tools are built to be opened up. API access will be announced here.