AgriVoice uses a WhatsApp voice note as the pilot interface. A cocoa farmer can ask in Asante Twi, hear a reviewed answer in Twi, or be connected to an extension officer when the system cannot safely answer.
At 6:40 one morning, Yaw is standing at the edge of his cocoa plot with a leaf between two fingers. It has dark spots he has not seen before. His phone is already in his hand because he used WhatsApp the night before to send a family voice note.
He does not open an unfamiliar app. He does not search through a menu, type a crop name, or work out which button describes his problem. He holds the WhatsApp microphone and asks his question in Twi.
The stakes can be real. If the advice is wrong, he could spend money on the wrong input, delay action on a spreading problem, or spray a chemical without the information needed to use it safely. A confident reply would be worse than silence.
The familiar microphone carries the question
For this pilot, the voice note is the front door. WhatsApp already provides the record button, playback controls, and conversation history. That matters when the goal is to learn whether a farmer can get useful guidance, rather than test whether they can learn a new interface.
Yaw sends his recording. AgriVoice transcribes the question and uses the meaning of it to select from reviewed advice blocks. The system does not write fresh agronomy guidance. It chooses content that has been prepared for the service, then returns it as speech.
That boundary matters most when the question touches pesticides. Advice on dosage, re-entry, and pre-harvest intervals cannot be guessed. When those details have not been verified, the safe response is to stop and escalate. The bottle stays closed until someone accountable can answer, as explored in Cocoa pesticide advice: Why AgriVoice Stops When Dosage Details Are Missing.
A reply must be understandable before it is useful
A few moments later, Yaw receives a voice reply in the same chat. He can replay it while looking at the affected leaves. He can share it with someone beside him. He can ask a follow-up without starting over.
The pilot still has hard gates before that moment can reach farmers. The reviewed advice needs spoken Asante Twi from a farming-aware translator. Chemical blocks need a Ghanaian agronomist’s approval where dosage and safety intervals are involved. A cocoa-sector partner must name the extension officer who receives questions that need a person.
The voice itself also needs scrutiny. Neuralis has an existing Asante Twi voice for the pilot, but it was trained on scripture recordings and can sound formal. Field users must be able to understand and accept it. A response delivered quickly in unfamiliar or unclear speech has failed its purpose.
Farmerline has reported that English, text-only advisory can exclude people with limited literacy, while local-language voice interaction can work for people who cannot read, write, or use smartphones. The practical lesson is smaller and more demanding: meet people in a channel they already use, then prove every part of the exchange works.
Silence and escalation are part of the interface
Yaw’s question may not match a reviewed answer. The recording may be unclear. It may contain a chemical question with missing verified details. In those cases, the system needs to say so plainly and pass the question to the named extension officer.
That is still an answer. It prevents the most damaging failure: a farmer receiving unreviewed instruction delivered with confidence.
Later that afternoon, Yaw listens again to the reply in his WhatsApp thread. He has one place to return to, one familiar microphone for the next question, and a clear route to a person if the advice is uncertain. The screen has done its job quietly. The conversation remains the interface.
Comments
No comments yet.