Twi, Ewe, and Ga in. English records out.
Nebula runs WhatsApp and voice in the customer's language and writes each interaction to your core system as a structured English record.
- 3 + English
- Twi, Ewe, Ga, and Ghanaian-accented English in one recognition pass
- 2 channels
- WhatsApp messages and voice calls
- 1 record
- One Interaction Record written per interaction, in your schema
- 0 external model calls
- Recognition, translation, routing, and record writing run inside your perimeter
THE UNIT OF OUTPUT
Intent, resolution, transcript, and translation, captured in your schema.
An Interaction Record captures one customer interaction at the moment it happens: intent, resolution, customer identifier, timestamp, languages spoken, the original transcript, and its English translation. Source audio is retained inside the perimeter and linked by reference.
Field names and structure are mapped to your existing schema during integration. Once records accumulate, contact volume by workflow, resolution rate by category, recurring complaint themes, and the language mix of your customer base become queryable reports.
{
"record_id": "IR-2026-0915-004817",
"timestamp": "2026-09-15T10:42:07Z",
"channel": "whatsapp_voice_note",
"languages_detected": ["tw", "en"],
"customer_ref": "ACC-****3391",
"workflow": "collections.payment_arrangement",
"intent": "request_instalment_plan",
"transcript_source": "Mepɛ sɛ metua me loan no kakra kakra. Can I pay 500 cedis this Friday?",
"transcript_en": "I want to pay my loan little by little. Can I pay 500 cedis this Friday?",
"resolution": "arrangement_proposed",
"arrangement": { "amount_ghs": 500, "first_due": "2026-09-18", "frequency": "weekly" },
"escalated": false,
"audio_ref": "perimeter://audio/2026/09/15/004817.opus",
"destination": "core_banking.collections_case"
}HOW A RECORD IS PRODUCED
Five steps from a voice note in Twi to a record in your system of record.
- 01
The customer speaks
A borrower sends a WhatsApp voice note in Twi, switching into English to name an amount and a date.
- 02
Nebula recognises it
Twi, Ewe, Ga, and Ghanaian-accented English are recognised in a single pass, with no language menu and no IVR tree.
- 03
Nebula routes it
The utterance is classified to a configured workflow; anything outside that set goes to a human operator with the transcript attached.
- 04
The customer gets an answer
The reply returns in the customer's language, on the same channel, in the same session.
- 05
The system gets a record
The Interaction Record lands in the core banking platform, CRM, or case management system, in the fields that system already expects.
WHAT AN INSTITUTION MODELS
Record sets follow the workflows an institution runs.
| Sector | Workflows recorded | Reportable once records accumulate |
|---|---|---|
| Retail and microfinance banks | Collections and payment arrangements, balance and transaction enquiries, disputed debits, branch and agent referrals, pre-application product questions | Recovery activity and product demand by language |
| Distribution utilities | Customer meter readings, billing disputes, fault and outage reports with location, reconnection requests | Fault history by area, queryable across languages |
| Public agencies | Citizen enquiries by programme, registration and eligibility intake, document requirement explanations, campaign follow-up questions | Programme coverage by language and district |
WHERE IT RUNS
Every processing stage deploys inside your infrastructure.
Customer audio is personal data. Many regulated institutions restrict it from leaving their infrastructure, by regulation, internal policy, or client contract. Nebula is designed around that constraint.
| Component | Runs in |
|---|---|
| Recognition and translation models | Client-controlled private cloud or on-site hardware |
| Routing layer | Client-controlled private cloud or on-site hardware |
| Record writer and API integrations | Client-controlled private cloud or on-site hardware |
| Source audio, transcripts, records | Retained inside the perimeter from receipt to deletion |
Once a message reaches the institution, audio, transcripts, and records stay inside its boundary until deletion.
Scope
Nebula handles
Speech recognition, translation, routing, customer response, and structured record output, currently across Twi, Ewe, and Ga.
Your institution keeps
Credit determinations, eligibility rulings, clinical judgements, and the core system of record. Nebula writes to that system and leaves its decisions where they are.
Additional languages are scoped per engagement.
Common questions
What is Nebula?
Nebula is AdwumaTech's platform for running customer channels in the languages an institution's customers speak. The current release provides speech recognition, machine translation, and message routing across Twi, Ewe, and Ga on WhatsApp and voice, and writes an Interaction Record to the institution's core systems for every interaction handled.
Which languages does Nebula support?
The current release supports Twi, Ewe, and Ga, plus Ghanaian-accented English, with code-switched speech handled in the same recognition pass. Further languages are added per engagement, scoped to the customer base the deployment serves.
Can Nebula run on-premise or in a private cloud?
Yes. Recognition, translation, routing, and record writing deploy in a client-controlled private cloud or on local hardware in the institution's own facility. Audio, transcripts, and records are retained inside the perimeter until deletion.
Does Nebula replace our existing systems?
No. Nebula delivers records through standard APIs into the core banking platform, CRM, or case management system already in place. The integration surface is the channel and the record format.
What is Nebula trained on?
A speech and text corpus collected in Ghana. Two components, mGhana-ST and UGSpeechData, are public on Hugging Face under MIT license; the remainder is proprietary and collected to the same documented standard.
How does an engagement start?
With an AI Opportunity Assessment. Teams that want an initial read first can take the Deployment Readiness Assessment, which scores a single use case across four pillars and returns a maturity tier before any build commitment.