A voice service can fail a farmer even when every word is technically in Twi, if the voice uses a formal register the farmer does not readily understand. Record real farmer questions and test comprehension before scale, because laboratory scores cannot show whether spoken advice lands in a working conversation.
In 1999, NASA lost the Mars Climate Orbiter as it approached Mars. The spacecraft had travelled for months, but its navigation calculations combined measurements in different unit systems. The result was a trajectory that brought it too close to the planet. NASA’s Mars Climate Orbiter Mishap Investigation Board documented the failure: information had been produced, passed along, and used, yet it did not mean the same thing at every point in the system.
That is the uncomfortable shape of a local-language voice service failure. A model can recognize speech. A reviewer can approve written Twi. A speech engine can produce audio. Then a farmer hears the reply and has to ask a neighbour what it meant, or acts on a phrase that sounded authoritative but unfamiliar.
For cocoa advice, that gap matters. A missed nuance about a diseased pod, a spraying precaution, or when to wait for a human response carries more weight than a polished demo.
Formal Twi can sound correct and still fail in conversation
Neuralis’s current Asante Twi voice was trained from scripture recordings. It can serve the AgriVoice pilot demo, but the source material gives the voice a formal character. A native speaker scored it at 55% on first hearing. That is useful evidence, because it describes the actual risk rather than hiding it behind a technical label.
A farmer does not need a voice that sounds impressive in isolation. They need one that can be understood while they are standing near cocoa trees, switching between Twi and Ghanaian English, and trying to decide what to do next.
The system also has to handle the language people already use. Farmers may say English words such as “fungicide” or “COCOBOD” inside a Twi question. Numbers, prices, percentages, phone numbers, and years are often spoken in Ghanaian English. A voice that treats these everyday patterns as edge cases can turn a practical answer into an extra burden.
That is why the first field recording matters so much. It reveals the distance between text that looks acceptable to a reviewer and speech that a person can follow on the first listen.
A recording tests the whole path, not one component
The AgriVoice pilot requires at least 20 real farmer questions to be recorded and transcribed before farmer exposure. Those recordings are needed to measure speech recognition on the kind of speech the service will actually receive.
Read speech is useful, but it is not the same as a voice note sent from a farm. Real questions may be brief, indirect, code-switched, interrupted, ambiguous, or shaped by local vocabulary. The speaker may describe what they see rather than use the disease name in a content block. A transcript can show whether the system heard the question. A follow-up comprehension check can show whether the farmer understood the reply.
Both measurements belong together.
For AgriVoice, the response path is deliberately constrained. The language model selects from reviewed answer blocks rather than writing agronomy advice. Missing or uncertain answers go to a named extension officer. That reduces one class of risk, but it does not remove the need to test the spoken response. A safe answer that cannot be understood has still failed the farmer.
This is also why chemical guidance remains withheld where dosage, re-entry, and pre-harvest intervals have not been verified by a Ghanaian agronomist. The service should escalate rather than fill silence with confident language. What Happens When a Cocoa Question Needs a Human Answer? explains why that handoff needs a real owner.
Intelligibility needs a pass or fail threshold
Technical speech metrics are valuable, but they answer a narrower question. Neuralis measured a mean character error rate of 0.300 on 25 held-out sentences for its final Asante Twi voice. The planned pilot scorecard adds a human measure: at least 80% of participants should report understanding the response.
That target is deliberately plain. It asks whether a farmer can use what they heard.
The pilot should capture more than a yes or no response. After a farmer hears an answer, ask them to explain the next action in their own words. Did they understand the advice? Did they hear a key term differently? Did they need the message replayed? Did the formal voice make them hesitate? If the answer concerns chemicals, did the system correctly escalate instead of offering an incomplete instruction?
A recording also makes correction possible. If consent and deletion behaviour have been explained and tested, the team can review recurring recognition or comprehension failures without treating farmer speech as an extractive dataset. What Happens When a Farmer Wants Their Voice Recording Deleted? covers that responsibility.
Scale should follow evidence from the farm
NASA’s Mars Climate Orbiter did not fail because people ignored data. It failed because a critical mismatch survived long enough to affect the mission. In a voice service, the equivalent mismatch can sit between formal audio and everyday understanding.
The practical next step is modest: collect consented farmer voice questions, transcribe them, run them through the full reply path, and check comprehension with the people who asked. Keep examples that fail. Group them by cause: recognition, code-switching, unfamiliar wording, voice quality, missing content, or an escalation that did not reach its owner.
A two-week pilot with 20 to 50 farmers can produce that evidence. It can show whether people return in week two, whether the answer arrives quickly enough, and whether the voice earns understanding before anyone tries to expand it.
Comments
No comments yet.