Listens
A speech recognition model (Whisper) transcribes each voice note within seconds, on our own servers.
audio 0:04
→ « nous sommes sept à la maison »
→ « nous sommes sept à la maison »
Artificial intelligence
AI that listens, understands, checks and learns.
In Chad, many people would rather speak than write. SahelVoice puts speech recognition to work for surveys, with a person to check whatever the machine is not sure it understood.
Understands
It pulls out the expected answer: a number said in words, a yes or a no, a choice from the list, a rating from 1 to 5.
« je donne quatre »
→ service_rating = 4
→ service_rating = 4
Checks
A spoken number is read back to the person to confirm. An answer it cannot understand is asked again once, then sent to your enumerators.
“You said: 7. Is that right?”
→ Yes ✓
→ Yes ✓
Learns
Every enumerator correction measures the real transcription accuracy. With separate consent from respondents, it helps train our models for the languages of Chad.
« cette » corrected to « sept »
→ error rate measured
→ error rate measured
We do not publish an accuracy figure until we have measured it on real answers from Chad. The dashboard computes that figure from your own enumerators' checks.
The platform
The whole survey cycle, in one place.
From importing the questionnaire to the final export, SahelVoice replaces paper forms and one-by-one phone calls.
Your questionnaire as it is
Upload the XLSForm file from KoBoToolbox or ODK. Simple skip logic is kept; anything WhatsApp cannot do is flagged clearly.
Invitations and QR codes
Invite a contact list, or put up a keyword and a QR code. The person writes in, and the survey starts.
Live results
Each answer reaches the dashboard the moment it is sent, with the audio ready to play.
Review queue
Your enumerators listen to voice notes one by one: correct, fix, or inaudible. The next one opens by itself.
Analysis-ready export
One row per person with the checked values, in Excel or CSV. Audio in a ZIP, named like the columns.
Team and roles
Administrators and enumerators each have their own access. Every organisation sees only its own surveys.
Human + AI
AI suggests. Your enumerators decide.
The machine does the repetitive work. The people who know the field have the final say on every voice answer.
- Voice notes to check wait in a queue, ready to play.
- A correction replaces the value in the export, and the original is kept.
- The Quality panel shows the real error rate and the share of answers corrected.
Post-distribution monitoring · examplelive
Responses48
Voice19
To check6
full_kit“yes we got everything, thank you”checked
household_size« nous sommes cette à la maison »fixed: 7
missing_item“the soap was missing”to play
Illustrative example, not real data. Answers are in French; shown translated.
Who it is for
For anyone who needs to hear from people.
NGOs and agencies
Post-distribution monitoring, beneficiary feedback, project evaluation. People who rarely write can answer out loud.
“Did you receive the full kit?”
Businesses and shops
Customer satisfaction, delivery feedback, service reviews. A QR code at the counter is enough to start.
“How would you rate our service?”
Training and associations
Feedback on a training, a workshop or a mentoring programme, with results visible live.
“What did you learn today?”
Languages
French today. The languages of Chad next.
Large speech models understand the languages of Chad poorly, for lack of data. We are building that data with communities, only with people's explicit consent.
French
Transcription and understanding
● availableChadian Arabic
Speech data collection
○ in preparationNgambay
Speech data collection
○ in preparationData protection
Respondents stay in control.
Consent first
Asked before the first question. Without consent, nothing is recorded and the person appears nowhere.
STOP means stop
Writing STOP ends all messages, from every organisation, for good.
Phone numbers hidden
Phone numbers are replaced by a code in exports and hidden in the dashboard.
Audio deleted on time
Voice notes are erased after the period the organisation sets. Reusing them for research needs a second, separate consent.
About
Built in Chad, for Chad.
Sakayo Toadoum Sari
Founder · AI researcher
Researcher in natural language processing for low-resource African languages. In Chad, almost everyone has WhatsApp, and many people would rather speak than write. SahelVoice turns those conversations into reliable data for the organisations that serve the population.
Pilots
Launch your first voice survey.
Are you an NGO, a business or an association working in Chad? The first pilots are free. We set up the survey with you.
- +235 66 04 90 94
- Head office
- N'Djamena, Chad