VoiceBridge — One-Page Project Brief
Project type: Non-profit, volunteer-run education technology project. Early prototype and pilot-preparation stage. Not yet a registered charity; we are seeking a local partner organisation in Africa for field execution and fiscal sponsorship. We sell nothing, have no commercial revenue, and do not handle donations ourselves.
Problem
273 million children are out of school worldwide (UNESCO). What they lack is rarely a textbook — it is someone asking a question and waiting for the answer. Strong existing programmes (solar tablets, mother-tongue courseware, broadcast and radio lessons) serve children who are in school. A literacy app quietly assumes school, a teacher and a starting level of reading. For a child outside the system, none of the three holds.
Approach
A low-cost AI voice device that talks to the cloud AI and to nobody else — no remote human on the line:
- A child speaks into a near-field microphone → cloud ASR + LLM + TTS → a spoken answer in their own language
- Shared hub first: one unit with a loud speaker, so everyone hears every answer and peer learning comes free. A centre needs only one or two units, fixed in place or carried to where children already gather.
- Handheld later: one per child, only after validation, enabling individual progress; numbered, stored at the centre, never taken home
- Topic whitelisting, sensitive-term filtering and time-window limits
- When confidence is low it does not guess: it asks the child to repeat, and otherwise hands over to the on-site coordinator
- No raw audio is stored — only counts, durations and topic categories
Audience
Africa (country to be decided), children aged 6–15 who are not in the school system. Pilot: one centre, one or two hub units, three months.
Technical requirement
One small VPS to host a bilingual project website, device dashboard and pilot data. Traffic is expected to be low.
| Item | Requirement |
|---|---|
| CPU / RAM | 1–2 vCPU / 1–2 GB |
| Storage | 20–40 GB SSD |
| Network | Dedicated IPv4, SSH access |
| TLS | Let's Encrypt support |
| Term | 12 months, renewable based on activity |
Why a donated VPS matters
The project has no revenue. Early funding goes to prototypes, speech-accuracy testing and local training. A donated server keeps limited money on the child-facing side.
What we have not yet proven
Speech recognition on children's speech — especially in low-resource languages — is far weaker than on adults: roughly 5% word error rate (WER) for adults, 15–21% for children aged 6–10, and higher in some low-resource settings. This is the core risk of the project. We will publish our own measured error rate before any pilot; if it misses the bar we do not proceed, and we say so publicly. One thing we do not leave to chance: recognition may be approximate, but the answer a child hears may not. The most common replies are pre-recorded by native speakers and reviewed one by one; only the long tail is synthesised. Nothing a child hears is generated on the fly. Because there is no human on the call, nobody online can correct a mis-recognised question — the on-site coordinator is the only human backstop, and is mandatory.
Milestones
| Phase | Timing | Goal | Gate |
|---|---|---|---|
| Phase 0 | Month 0–1 | Site live, partner conversations | At least one local organisation in substantive discussion |
| Phase 1 | Month 1–2 | 1–2 hubs + speech test with 20 children | Intent-routing accuracy ≥ 80%, and ≥ 10 points better near-field than far-field |
| Phase 2 | Month 2–5 | One centre, 1–2 hubs, 3 months | Mid-term: ≥ 4 sessions per week |
| Phase 3 | Month 5–6 | Cost accounting, iteration list | All KPIs met before handhelds are considered |
Sustainability and giving back
- Sponsor acknowledged in the site footer
- Firmware, admin backend and documentation open-sourced where privacy and safety allow
- Regular public updates on pilot progress and a spending summary
- Pilot findings — including what did not work — published for reuse
Contact
Name: ______ Email: ______ Website: ______ GitHub: ______
(Available on request: one-page brief, budget, milestones, privacy and safeguarding policy drafts, prototype photos and measured accuracy data.)