Outbound AI voice agent for cold calls in Nairobi by Artūrs BidiņšOutbound AI voice agent for cold calls in Nairobi by Artūrs Bidiņš

Outbound AI voice agent for cold calls in Nairobi

Artūrs Bidiņš

Artūrs Bidiņš

Most cold calls in Nairobi fail before the pitch, because the voice on the line sounds like it is calling from another continent.
I built an outbound voice agent for a security company in Nairobi that calls prospects in English with an East-African voice, and follows the conversation into Swahili when the prospect goes there.

Finding the voice

Most text-to-speech libraries have no African voices at all. I worked through nine providers - ElevenLabs, Cartesia, Deepgram, Hume, Rime, MiniMax, Inworld, WellSaid and Azure - auditioning everything each of them had. A British voice read as an outsider. Nigerian and South African voices were the wrong region for the people being called. A Swahili voice was unusable over a phone line. A Tanzanian East-African English voice won, and the client approved it after live test calls.

Making it survive a real conversation

The first version talked over people. Fixing that meant waiting through "eee" and "emm" instead of treating every pause as the end of a turn, ignoring the short words someone drops in just to show they are listening, and picking an interrupted sentence back up instead of abandoning it. Each turn is capped at one question and under twenty-five words, because an agent that asks two things at once loses the caller.

What happens on a good call

The agent opens, gets past "who is this", works out whether there is a real need, and books a short follow-up. Only after a clear yes does it forward the lead - and it reads the phone number and email back digit by digit first, because a cold call is worth nothing if the callback number is wrong. The lead then arrives in the sales inbox through an n8n workflow with name, role, company and what the prospect actually said they needed.

Keeping the client's data separate

Access to the voice platform is account-wide, which would have meant the client seeing every other agent in my workspace. I set up a separate workspace for them and cloned the agent into it, so they see their agent and nothing else.

Reliability work

Model temperature lowered for consistency, fallbacks configured for both the model and the voice, background noise suppression enabled, call duration and silence limits set, and digit-by-digit read-back added before any lead is sent. The things that break a live phone agent are rarely the interesting parts.
Like this project

Posted Sep 1, 2026

An outbound voice agent that cold-calls in English with an East-African voice, follows the prospect into Swahili, qualifies the need and books a follow-up.

Likes

0

Views

0

Timeline

Jul 22, 2026 - Jul 30, 2026