Telroi Research

Teaching machines
to hear growing markets

Speech models built from real calls — starting in Nigeria: Hausa, Yoruba, Igbo and Pidgin.

Read the research →
How we build it

Real voice.
Real consent.

We don't generate training data. Every hour comes from real conversations across our own network and partner MVNOs — collected with consent, de-identified, and labeled by people who speak the language.

Real human voice, never synthetic
Consented & de-identified
Collected across our network + partner MVNOs
SPEECH SEGMENT · 00:07LISTENING
“Ẹ káàrọ̀, mo fẹ́ mọ̀ bóyá ẹ ṣílẹ̀kùn lọ́jọ́ Sunday.”
“Good morning, I want to know if you're open on Sunday.”
ConsentedDe-identified
yo-NG · gold
Progress, in the open

Where we are today.

1,243hrs
Real labeled speech
4
Languages
38,920
Gold segments
14.2%
Best WER · Yoruba

Research & papers

Published as we go

Building the voice layer for growing markets.

We publish our datasets, methods and progress in the open — and work with partners who want speech that understands their customers.

Talk to sales →