The Cocoa Translation That Turned “Kokoo” Into “Chicken,” and What It Could Cost

A service reaches cocoa-farming households only when people can understand and use the guidance it carries. For Asante-Twi-speaking farmers, reviewed advice delivered as spoken Asante Twi through WhatsApp can close the gap between having access to information and comprehending it well enough to act safely.

In April 1970, the Apollo 13 crew had carbon dioxide filters aboard their damaged spacecraft. The problem was fit: square command-module cartridges could not connect to the lunar module’s round openings.

In Houston, NASA engineer Ed Smylie and his team worked with materials available to the astronauts, including plastic bags, cardboard and tape. Their improvised adapter allowed the filters to do their job. NASA’s account of the Apollo 13 mission documents the crisis and the solution.

The filters were present. Until the crew could use them in their actual environment, that availability offered little protection.

A programme manager planning services for cocoa-farming households faces the same practical test. A message can sit in a database, appear on a screen or arrive on a farmer’s phone. That proves delivery. It does not prove understanding.

The last mile ends with comprehension

Suppose reviewed cocoa guidance is available in English. A farmer can receive the message on WhatsApp, yet still struggle with a technical phrase, an unfamiliar disease name or a long written explanation. Another person may have to read it aloud. The meaning may change as it passes from one person to another.

Even translated text can miss the mark. Formal written Twi may differ from the Asante Twi farmers use in conversation. Agricultural vocabulary also carries risk. In Neuralis testing, machine translation changed “kokoo,” meaning cocoa, into “chicken.” The result was fluent enough to sound credible and wrong enough to undermine the entire service.

Spoken Asante Twi changes the point of access. The farmer can ask a question by voice and hear the reviewed response in a familiar language. WhatsApp provides a channel people can use on a phone without learning a separate application.

That still does not make every answer safe. The content must fit the question, the speech must be understandable, and uncertain cases must reach a person who can respond.

Reviewed guidance sets a safety boundary

AgriVoice is designed around reviewed content blocks. The language model selects from those blocks; it does not compose agronomy advice from scratch.

That distinction matters most when a question involves pesticides. Product names alone are insufficient. Dosage, re-entry periods and pre-harvest intervals can determine whether workers or crops are exposed to avoidable harm. When those details have not been verified by a Ghanaian agronomist, the system withholds the chemical guidance.

A useful voice service must therefore be able to say, in effect, “This needs a person.” The programme also needs a named extension officer who owns the escalation queue and answers within an agreed period. What Happens When Reviewed Twi Cocoa Advice Lacks Safe Spraying Details? examines why translated advice still requires complete safety details.

This approach may produce fewer immediate answers than unrestricted generation. It gives the programme manager a clearer promise: the system will speak reviewed guidance when there is a safe match and escalate when there is not.

Measure what happens after delivery

A WhatsApp delivery receipt cannot show whether a farmer understood the response. Neither can a count of audio files sent.

A practical pilot should measure the outcome closer to the household. Neuralis plans a two-week pilot with 20 to 50 farmers and a cocoa-sector partner. Its comprehension threshold is concrete: at least 80% of participants should report understanding the response. The pilot will also look for at least 30% of activated farmers returning with another question in the second week.

Those measures expose different failures. Low comprehension may point to translation, pronunciation, accent or an explanation that is too formal. Low repeat use may mean the service solved no recurring problem, even if the first answer sounded clear. Repeated confident errors require stopping or redesigning the workflow.

Field speech matters too. A speech recognizer that performs well on clean, read sentences may struggle with real farmer questions, code-switching and background noise. At least 20 recorded and transcribed farmer questions are required before exposure so the programme can measure recognition on the voices it intends to serve.

Build the service around the household

The first operational step is to choose one narrow set of reviewed cocoa questions. Have a farming-aware Asante Twi translator render each response as speech a farmer would naturally hear, then ask a Ghanaian agronomist to clear any chemical instructions.

Next, test the entire path: a real voice question through WhatsApp, transcription, constrained content selection, spoken response and human escalation. Test consent and deletion behaviour as part of the same exercise.

Finally, listen with farmers. Ask what they understood and what they would do next. A correct answer that leaves the listener unsure has failed its purpose.

Apollo 13’s cartridges became useful only after Smylie’s team made them fit the crew’s real conditions. Cocoa guidance follows the same rule. The programme manager’s job ends after reviewed knowledge fits the language, channel and safety needs of the farming household.