When a spraying team is already unloading at a cocoa farm, AgriVoice must either select a reviewed answer or send the question to a named extension officer. It must never fill a gap with confident pesticide advice.
A farmer holds down the WhatsApp microphone while people wait nearby. The voice note moves between Twi and English: a product name, a cocoa disease term, a question about spraying today, perhaps a harvest date. The crew has the bottle. The question is urgent. That urgency cannot make an unverified dose, re-entry interval, or pre-harvest interval safe.
The decision has to survive the field
This is the condition AgriVoice is being built for: speech comes in as farmers actually speak, including code-switched terms and imperfect audio. The system’s job is narrow on purpose. It recognizes the question, selects from reviewed guidance blocks, returns a spoken response in Asante Twi, or escalates when the right reviewed answer is missing or uncertain.
A fast answer only helps when it is safe to act on. A question about leaf symptoms may match approved guidance. A question that depends on a chemical dosage or a harvest interval needs verified details. If those details are not in the reviewed content, the farmer should hear the safe next step and the escalation should reach the extension officer responsible for the pilot.
That is why the pilot gates matter before farmer exposure. The cocoa-sector partner must own recruitment. A named extension officer must receive escalations. The spoken Twi guidance needs review by a farming-aware translator. Chemical blocks remain withheld until a Ghanaian agronomist verifies the information that could change how someone sprays, re-enters a field, or harvests.
Apollo 13 had the parts, but not the fit
In 1970, Apollo 13’s crew faced a rising carbon dioxide problem after the spacecraft was crippled on its way to the Moon. The lunar module had square lithium hydroxide canisters; the command module had round ones. The available equipment could remove carbon dioxide, but it could not connect without an improvised adapter.
At NASA’s Manned Spacecraft Center in Houston, engineer Ed Smylie led the team that worked on the solution. The crew ultimately used a device assembled from materials already aboard, including plastic bags, cardboard, tape, and a flight-plan cover. The outcome had been uncertain. The solution had to work with the actual constraints of the spacecraft, using instructions the astronauts could carry out.
NASA’s Apollo 13 Flight Journal documents the mission and the effort to adapt the square canisters for the lunar module system.
The parallel for a cocoa farm is straightforward. A language model may recognize familiar words, but recognition alone does not make an instruction usable or safe. The answer has to fit the farmer’s exact situation, the available reviewed content, and the safety limits around a pesticide question. When those pieces do not fit, escalation is the correct output.
Reviewed guidance creates a boundary people can trust
AgriVoice is designed so the model selects reviewed block IDs. It does not author agronomy advice. That distinction keeps the system from turning an uncertain voice note into a polished instruction that sounds authoritative.
The boundary is especially important because translation can quietly change a domain noun. Neuralis has already seen `kokoo`, meaning cocoa, return as “chicken” in machine translation. A fluent sentence can still carry the wrong meaning. For a farmer deciding what to put in a sprayer, that failure is serious.
The pilot will test roughly 40 real Twi questions across common questions, ambiguity, pesticide topics, out-of-domain requests, code-switching, noisy speech, and questions that need more than one guidance block. The same cases will be tested again using real ASR transcripts. The team will measure whether the selected set is correct, whether the system adds an unsafe extra block, whether it escalates appropriately, and how long the response takes.
This is practical safety work, not a performance demo. A response that reaches a farmer quickly but directs them wrongly has failed. Cocoa pesticide advice: Why AgriVoice Stops When Dosage Details Are Missing explains why some questions must remain unanswered until an expert verifies the missing facts.
The safe response must arrive while the decision still matters
The two-week pilot aims to measure a successful answer rate of at least 70%, counting questions that are answered correctly or correctly escalated. It also requires zero unsafe pesticide answers, at least 80% reported comprehension, and a median end-to-end response under eight seconds, excluding human escalation.
Those measures reflect the moment at the farm. A farmer may need immediate guidance on what can safely happen next: pause spraying, keep the bottle closed, use protective equipment, or wait for the extension officer. A useful system gives that next step plainly and records the unresolved question for a person who can answer it responsibly.
Apollo 13’s adapter succeeded because it joined a specific square part to a specific round opening under pressure. AgriVoice needs the same discipline. It should connect a real voice question to guidance that has been reviewed for that exact decision. Where the fit is missing, the system must say so before the spraying starts.
Comments
No comments yet.