More voice-model training is not Neuralis’s next milestone because AgriVoice needs reviewed Twi advice, consented farmer speech, and a safe route to a human before another language model can help a farmer. We paused new-language work to prepare one two-week pilot for 20 to 50 Asante-Twi-speaking cocoa farmers with a cocoa-sector partner.
At 6:40 one morning, an illustrative farmer named Kofi stands at the edge of his cocoa plot with a pod in one hand and his phone in the other. Rain has left the leaves wet. He records a Twi voice note about marks spreading across the pods, then waits for an answer he can understand before deciding what to do.
The bad ending is clear. If the service confidently gives him the wrong disease advice, he could treat the wrong problem or delay calling the extension officer until the affected pods have spread further. A better-trained voice would not solve that. Clear audio only makes a wrong answer easier to hear.
The first job is to give the system safe words to say
AgriVoice is designed to turn a farmer’s voice note into a selection from reviewed guidance, then speak that guidance in Twi. The language model selects a reviewed content block. It does not write agronomy advice from scratch.
That boundary matters because fluent wording can hide a serious mistake. Early translation work produced a warning that was hard to ignore: “kokoo,” meaning cocoa, came back as “chicken.” The sentence could sound polished and still send the conversation in the wrong direction.
Before farmers use the pilot, the approved guidance needs spoken Asante Twi from a farming-aware translator. The content has to sound natural when read aloud, retain the practical details, and flag anything that may be wrong or unclear in the English source. Three chemical-related blocks remain held because dosage, re-entry, and pre-harvest intervals have not been verified by a Ghanaian agronomist.
That restraint is part of the product. When the needed detail is uncertain, AgriVoice should escalate the question to a named person rather than fill the gap with a plausible answer. What Happens When Spraying Starts Before Pesticide Details Are Verified? explores why that pause protects the farmer.
For Kofi, the turn comes when the service can safely recognize that his question needs a reviewed disease-response block or a human follow-up. The useful outcome is not a long message. It is a clear next step in his language, with no invented spray instruction.
Farmer speech must shape the next speech-recognition decision
Neuralis already has a Twi voice trained from scripture recordings. It can support the pilot demo, but it sounds formal, and a native speaker scored it 55% on first hearing. More training on the same material would not create the conversational rhythm of a cocoa farmer leaving a voice note beside a wet plot.
The next evidence must come from consented farmer speech. The pilot entry gate requires at least 20 real farmer questions to be recorded and transcribed for field speech-recognition measurement. That work tests the conditions that matter: agricultural vocabulary, code-switched English, background sound, differing speaking styles, and the way a question changes when someone is worried and speaking quickly.
Consent is part of the test, not paperwork after the fact. Farmers need to know what is recorded, how it is used, and how deletion works. What Happens When a Farmer Wants Their Voice Recording Deleted? sets out why a deletion request has to reach the system’s actual data path.
Kofi’s voice note therefore has value beyond one reply. With his consent, an accurate transcription can reveal where speech recognition missed a crop term, confused a code-switched phrase, or failed to understand him at all. That finding points to a real improvement. Another round of generic model training does not.
A pilot measures safety, comprehension, and return use together
Neuralis is prioritizing one narrow proof: can a Ghanaian-language voice workflow deliver useful, safe answers through an institution that can recruit people and own human escalations?
The two-week AgriVoice pilot has concrete thresholds. At least 70% of questions should be answered or correctly escalated. No unsafe pesticide answer should reach a farmer. At least 80% of participants should report understanding the response. At least 30% of activated farmers should ask another question in week two. The median response, excluding a human escalation, should stay below eight seconds.
Each measure protects against a different kind of self-deception. A fast response does not matter if Kofi cannot understand it. A response that sounds helpful does not count if it gives unsafe pesticide guidance. A first voice note does not prove lasting value if nobody returns once the novelty has passed.
The named extension officer matters here. When the system cannot safely match a reviewed block, the question needs an accountable person and a response path that works in less than one working day. An unanswered queue is not an edge case. It is part of the farmer’s experience.
The next milestone is evidence a partner can act on
New languages remain important to Neuralis. They are simply not the work that most reduces risk today. Akuapem and Ewe voice results did not clear the release gate, and more single-speaker training does not outrank partner recruitment, reviewed content, field evaluation, and a working WhatsApp path for this pilot.
Kofi returns to his plot later with a response he can follow, or with confirmation that an extension officer will take the uncertain question. The phone is no longer waiting silently in his hand. That is the milestone worth measuring before Neuralis asks another voice model to learn more.
Comments
No comments yet.