The numbers

UNESCO’s most recent adult literacy estimate is roughly 754 million adults cannot read a short simple sentence. The ITU’s 2024 "Facts and Figures" report puts the global offline population at roughly one-third of the world — about 2.6 billion people with no regular internet access. GSMA Intelligence tracks smart-feature phones (feature phones with limited data) growing faster than smartphones in sub-Saharan Africa.

The overlap of these populations — low literacy, low bandwidth, feature-phone primary device — is where an app-first service has no traction. It is also where voice, as the natural first interface, wins by default.

Why app-first doesn’t reach these users

Four compounding constraints. Bandwidth is expensive (2G/3G dominant or spotty 4G). Device storage is small (feature phones and low-end Android). Literacy is not universal (reading long-form text in a colonial language is exclusionary). Onboarding via typed form fields is a non-starter.

The net result: every "tech for Africa / Asia / Latin America" startup that ships an English-first smartphone app sees a 90% churn curve against a 10% urban-elite retention curve. The addressable market looks huge on paper; the deliverable market is small.

What the voice-first pattern looks like

In Kenya, M-Pesa (launched by Safaricom in 2007) reached adoption first through USSD — not an app. USSD is ugly and limited, but it worked because it treated the common denominator (the feature phone) as a first-class citizen. That lesson generalises.

In India, Jio’s conversational voice launch surface demonstrated that, for a large share of the user base, the primary interaction mode is voice input to a cloud service rather than a typed query. Similar patterns appear in Bangladesh (bKash, Nagad), Egypt (Fawry), and the Philippines (GCash voice support).

The AI shift that makes this newly feasible

Speech-to-text quality has passed a threshold. OpenAI’s Whisper (2022, with multilingual extensions through 2024) is usable in dozens of languages. Open-source models (the MMS project, SeamlessM4T) have brought competent transcription to even more. Text-to-speech has followed a similar trajectory.

For LLMs, the jump is even larger. A 2019 voice-commerce system had to hand-craft intents. A 2026 one can parse free-form speech into structured action without a grammar. This is a genuine step-change.

What no-one has done yet

Pair the language-inclusive voice stack with actual commercial supply. Most voice-first products are media / entertainment (Spotify voice control, Alexa skills). Very few are real-world commerce (book a doctor, pay a bill, order food). That is the gap GeraVoice addresses — because the rest of the Gera portfolio already operates on the supply side.

The unlock for Gera

Every Gera vertical — GeraClinic, GeraEats, GeraHome, GeraCash, GeraRide, GeraLearn — has an API. GeraVoice turns those APIs into a voice interface. A user who cannot read an app menu can still book a doctor, because the interaction is spoken.

What voice-first does not solve on its own

Fraud. Voice-based fraud is very real; we will write separately about how we intend to resist it.
Consent. Recording consent by voice is harder to get right than consent by click.
Accent robustness. ASR degrades on accents underrepresented in training data; this is a live research problem.

Why now, why us

The STT quality is here. The LLMs are here. The commercial supply is here (the Gera portfolio). The populations are ready. What hasn’t shipped yet is the glue. That is GeraVoice.