AethexAI Raises $3M to Build Voice AI for Africa and the Middle East

A Goldman Sachs and Meta alumni duo just closed a $3M pre-seed to build voice AI that actually works in Africa, trained on call-centre recordings, radio audio, and handling code-switching dialects that Silicon Valley models choke on.

Ask any voice AI built in San Francisco to follow a customer call in Lagos, where a single sentence might slide from English into Yoruba into Pidgin and back, and watch it stumble. Add a noisy line, a local telephony network, and a model hosted on GPUs an ocean away, and the “future of customer service” starts sounding like an expensive answering machine.

AethexAI thinks it has the fix. The startup has emerged from stealth with a $3 million pre-seed round to build voice AI infrastructure designed from the ground up for Africa and the Middle East, a market of roughly 1.5 billion people where the telephone, not the chatbot, remains the front door of business.

The round was led by 4DX Ventures, with participation from Enza Capital, Dorm Room Fund, Mojo Ventures and the Stanford GSB ’26 Fund. The angel list reads like a deliberate signal of intent: Stanford faculty, telecom executives, and AI researchers from Anthropic.

The founders who heard the problem first-hand

AethexAI is the work of Mariama Diallo and Ayooluwa Odemuyiwa, a pairing of high finance and hard engineering. Diallo cut her teeth in investment banking at Goldman Sachs before becoming the first product and growth hire at YC-backed Model ML. Odemuyiwa studied computer science at Caltech, built engineering systems in aerospace and at Meta, then passed through Stanford’s Graduate School of Business.

The two founded the company in 2025 after spending time on the ground with businesses across Africa and the Middle East. What they saw was unambiguous: enterprises had already experimented with voice AI, and most of it had failed in production. Existing tools buckled under unreliable connectivity, fragmented telephony, high pricing and poor handling of local speech, often ending up more expensive than the human agents they were meant to replace. The efficiency gains everyone promised were sitting on the table, untouched. So they quit their jobs and started building.

Kora: small models, local ears

Rather than wrapping a product around someone else’s large language model, AethexAI made the harder, more contrarian bet: build the stack yourself. The result is Kora 1, a proprietary suite of speech models ranging from 300 million to 1.7 billion parameters, deliberately small, deliberately fast.

The training data is where the localisation gets real. Kora learned from licensed and anonymised call-centre recordings, audio from radio stations and regional content platforms, with annotation work done by university students. The models are engineered for the conditions Silicon Valley rarely thinks about: noisy lines, heavy accents, informal speech, and the constant code-switching between local varieties of English, French and Arabic that defines how people actually talk across the two regions. The reported result: sub-500-millisecond turn-taking — fast enough to feel like a conversation, not a queue.

The logic behind going small is brutally practical. Large models hosted on high-end GPU clusters in North America or Europe are often too slow and too expensive for real-time calls running over African and Middle Eastern telephony. Compact, self-hosted, market-localised models flip that equation, and give AethexAI a fundamentally different playbook from global voice giants like ElevenLabs, Deepgram, Sierra and Cognigy.

Already on the line

This is not a demo in search of a customer. AethexAI says it already handles more than 17,000 calls a day, targeting the high-volume enterprise workflows where voice still rules in these markets: customer service, debt collection, customer activation and KYC verification. The platform ships as a no-code interface plus APIs and SDKs, wired directly into managed telephony, orchestration and existing enterprise workflows, and the company is pairing it with telecom partnerships and forward-deployed engineers embedded with clients.

The fresh capital will fund enterprise rollouts, grow the engineering and go-to-market teams, and deepen coverage across key regional markets.

The bottom line

The deeper story here is a thesis: voice AI for emerging markets cannot be a hand-me-down from the West. Across Africa and the Middle East, voice is still the primary channel for banking, telecoms and customer support, but the infrastructure, the languages and the economics are different. AethexAI is betting that the winners in this market won’t be the labs with the biggest models, but the builders with the best ears.

If a model trained on Cairo call centres and Dakar radio can out-listen a trillion-parameter giant on a crackly line in Nairobi, the next chapter of voice AI won’t be written in Silicon Valley. It will be answered, in under half a second, somewhere between Casablanca and Johannesburg.

© 2026 Tropics Media Group. Original content distributed by Tropics PressRoom. All Rights Reserved.