Kojo’s cocoa becomes chicken. Wrong advice could cost him.

AgriVoice is designed to select a reviewed Twi answer, then speak it, because a fluent translation can replace a cocoa term with the wrong word without warning. Agronomy advice needs a controlled answer path: reviewed content for questions the system recognizes, and a human escalation when it does not.

A word changes the decision on the farm

Consider Kojo, an illustrative cocoa farmer in the Ashanti Region, standing at the edge of a plot late in the afternoon with a leaf folded in his hand. Brown patches have spread across a few pods, and he wants to know what to do before the next visit to the farm.

He sends a voice note in Twi. Somewhere in a translation step, the word for cocoa returns as “chicken.”

The sentence may still sound smooth. That is what makes the error dangerous. A system that confuses the crop can select advice from the wrong domain, or produce an answer that sounds certain while missing the farmer’s actual problem.

Kojo has no reason to inspect the hidden translation. He hears a confident voice and has a decision to make. If the advice points him toward the wrong action, the possible outcome is not a bad search result. He could spend money on an unsuitable input, delay a response to a disease, or apply a chemical without the information needed to use it safely.

That is why the first rule for AgriVoice is simple: the language model does not write agronomy advice.

Reviewed answers create a safer boundary

AgriVoice is being prepared for a two-week pilot with 20 to 50 Asante-Twi-speaking cocoa farmers and a cocoa-sector partner. The intended workflow is deliberately narrow.

A farmer asks a question by voice. Speech recognition turns it into text. The system then selects from approved content blocks that have been translated into spoken Asante Twi and reviewed by a farming-aware translator. The selected answer is spoken back. If the question is unclear, outside those blocks, or involves information that has not been cleared, the question goes to a named extension officer.

The language model’s job is selection. It chooses a reviewed block identifier. It does not fill gaps with polished prose, calculate a dose, or improvise a pesticide recommendation.

That boundary matters because cocoa advice carries details that generic language systems handle poorly. A farmer may use a local crop name, switch between Twi and English, describe a symptom indirectly, or send audio with background noise. Each step creates room for a small linguistic error to become a practical mistake.

Reviewed blocks also make it possible to inspect what the system actually said. A pilot team can compare a farmer’s question, the selected block, the spoken reply, and any escalation. With generated agronomy, the answer itself changes from one request to the next. Reviewing failures becomes harder, and confident wording can hide uncertain reasoning.

For related pilot safeguards, see Cocoa Farm Voice Service Testing: When Safe Escalation Matters More Than a Lab Score.

Chemical questions stay open until an expert clears them

Some questions deserve a refusal more than a quick reply. AgriVoice currently holds three chemical content blocks because verified dosage, re-entry intervals, and pre-harvest intervals are missing. The system should not guess those details.

Back in Kojo’s plot, imagine his next voice note asks whether he can spray and return to the field the following morning. A generated answer could make that sound easy. A reviewed system has a safer option: explain that the answer needs an extension officer, preserve the question, and send it to the person responsible for follow-up.

The named person matters as much as the software. An escalation queue without an accountable extension officer leaves the farmer waiting with the same decision and less time to make it. The pilot scorecard therefore treats a working escalation path as a core measure, alongside successful answers, comprehension, response time, repeat use, and zero unsafe pesticide answers.

A safe answer is sometimes a direct spoken instruction. Sometimes it is a clear handoff. Both need to work in the language the farmer uses.

Translation must earn its place in the workflow

The cocoa-to-chicken error exposed a broader lesson: translation can be useful without being trusted blindly. Neuralis is testing whether direct Twi selection can match or beat a translation-first path on real farmer questions, including ambiguous, code-switched, pesticide-related, and speech-recognition-corrupted cases.

The test is not only about accuracy. It also checks unsafe extra selections, correct escalation, latency, and cost. A direct Twi path would remove one opportunity for a domain noun to change meaning. It still has to prove that it can select the right reviewed answer without creating new safety failures.

Kojo’s useful outcome is modest and concrete. His question reaches a system that recognizes its limits. If it can match his concern to reviewed Twi content, he hears advice prepared for cocoa farmers. If it cannot, the question stays open for a human instead of being turned into a confident guess.