Trusted Voice Research Infrastructure for NGOs, Governments & Global Development Partners
Delivery

What happens between a research question and a finding somebody can act on.

Eight stages. Two of them are the ones most descriptions of research software skip — deciding who a study reaches, and getting the work back from devices that were nowhere near a network when it was collected. In the places this platform is built for, those two take the longest.

The eight stages

Whether your team runs these or we run them for you is the question the engagement models answer. The stages themselves do not change.

1 · Design

Before anything is asked

The research question becomes an instrument: the questions, their order and their logic, the languages the instrument will be administered in, and the consent statement a respondent hears before any of it. The design is registered, so the analysis later runs against what was actually planned rather than what was remembered.

Research designs supported
Structured survey, facility assessment, citizen feedback and qualitative work run as live designs. Mixed methods, observation and longitudinal panels are supported on the same structures. The seven registered modes.
Qualitative formats
Key informant interviews, in-depth interviews and focus group discussions are recorded as formats within qualitative work, with participant counts — not as separate research modes.

2 · Field preparation

Who and where

Contacts are loaded — from a spreadsheet, a workbook or typed in one at a time — and normalised into one unambiguous form so a number is never dialled twice or dialled wrong. An audience is then a rule over those people, and its size is counted from your own respondents when you ask for it.

Consent is not assumed
A contact can only be dialled where consent was recorded. A blank consent column is treated as unknown, never as permission, and somebody who has asked not to be contacted stays that way through a re-import.
Enumerators and assignments
Field staff receive assignments with their own area and due dates, and a supervisor can see what has been given out and what is outstanding.
Offline packages
Where a team is going somewhere without a network, the instrument and its language pack are prepared on the device before they leave.

3 · Collection

Five routes

A person replies however suits them. A web link and the offline field application work wherever the platform is deployed. Phone, WhatsApp and SMS are connected once a communications provider is configured for your organisation — one account covers all three.

Voice, in any registered language
Recording someone's voice does not depend on which language they speak, so a respondent can answer aloud in any of the 57 African languages the platform registers, and the recording is kept as the evidence of what they said.
What that does not mean
Capturing speech is not understanding it. Automatic transcription reaches 15 of those languages on the published list of the speech provider we route to, and none of the 57 has been accuracy-validated by VoiceInsights against African field audio. The registry, language by language.

4 · Getting the work back

Often the longest

Work collected with no connection sits on the device until it is confirmed received. An interrupted interview resumes at the question it reached. A device that is running out of room says so, and refuses a new interview rather than silently failing to save one. If a confirmation is lost on the way back, the device sends again and the server recognises what it already holds instead of storing it twice.

A supervisor can see what has not arrived
Devices report the shape of their queue — how much is waiting, how long it has been waiting, and whether the last attempt was refused.
And what silence means
Those figures come from devices that reached the server. A device with no connection cannot report that it has no connection, so an absent device is unknown rather than healthy — and the platform says so rather than showing a reassuring zero.

5 · Evidence management

The record

Every response carries its consent status, its origin and its lineage. The original audio is retained — including a recording with no discernible speech, which is stored as evidence rather than discarded, because a refusal to answer is a finding and the recording is the proof of what happened.

6 · Quality control

A person decides

Responses are checked for completeness, duration, consent validity and duplication, and flagged where something does not hold. Double entry compares two independent records of the same interview. A supervisor decides; the platform records who decided and when.

Where AI is involved
Analysis assistance is separated from observation and from human interpretation, and is presented as suggestion rather than conclusion. Responsible AI.

7 · Analysis

Against the design

Descriptive results, frequencies and dispersion, with the cautions that belong to them. Qualitative coding keeps themes, contradictions and minority views rather than averaging them away. Findings reference registered evidence, so a claim can be traced back to the responses behind it.

8 · Reporting and decision support

Five formats

A governed publication is produced in the formats an institution actually circulates, and a publication whose evidence cannot be resolved is refused rather than produced.

PDF
The executive report, for circulation and for the record.
DOCX
An editable working document, for a team that will add its own programme context.
PPTX
A presentation, for the meeting where the decision is taken.
XLSX
A statistical workbook, for an analyst who wants the figures behind the finding.
CSV
The governed dataset export, for a data team or a donor's own pipeline.
Who can export what
Export is a permission, not a button everyone has. A role that may read a publication is not automatically a role that may compose or export one.

What this process does not do

Worth stating plainly, because each of these is something research software is routinely assumed to do.

It does not translate automatically
Translation is a governed artefact with a recorded reviewer, not something generated in the pipeline.
It does not detect a language for you
The platform performs no automatic language detection. The collection language is a decision that is recorded.
It does not identify who spoke in a group discussion
Focus group audio is preserved and attributed at the level a researcher can defend, not split by speaker automatically.
It does not let an operator message a respondent back
Collection is not a conversation channel. There is no reply surface, and there is no cross-channel message thread.
It does not decide anything on its own
Quality is scored; a person adjudicates. Safeguarding indicators, where configured, are raised for trained human review and replace no protocol or referral pathway.

Start a study

Tell us the question and the context. The proposal comes back with the design, the collection routes that suit it, the languages, the sample and the timeline.