Kofi's voice note waits in silence. The farming decision cannot wait.

AgriVoice will measure the full time from a farmer sending a WhatsApp voice note to receiving a spoken reply, with a pilot target of under eight seconds for automated answers. It will reduce delay by timing each step and improving the slowest one, while keeping answers limited to reviewed content and sending uncertain or unsafe questions to a named extension officer.

At 6:12 a.m., Kofi stands beside a cocoa plot outside Kumasi with a cut pod in one hand and his phone in the other. He has recorded a question in Asante Twi about marks spreading across pods near the edge of the farm. Rain is gathering over the trees, and he needs to decide whether to spend the morning checking more trees or leave for another job.

The voice note is understood correctly. Then the phone stays quiet.

A long pause changes the exchange. Kofi cannot tell whether help is coming, whether the message failed, or whether he should act on his own. If the answer arrives after he has left the farm, the useful moment has passed. If the question concerns a spray product and the system fills the silence with a confident guess, the risk is worse than delay.

AgriVoice has to earn speed without taking shortcuts around that second problem.

Measure the whole wait, then find the slow step

“Response time” can hide several different waits. A farmer records audio, WhatsApp delivers it, AgriVoice retrieves the media, speech recognition turns it into text, the selection system finds a reviewed answer block or chooses escalation, speech is generated, and the reply travels back to the phone.

The pilot will record the timing of those stages for every completed question. The central performance measure is median end-to-end response time under eight seconds, excluding the time a human needs to answer an escalation. Median matters because one unusually slow connection should not make an otherwise usable exchange look broken. Individual slow cases still matter, especially when they repeat.

That record gives the team something concrete to fix. If audio retrieval is slow, changing answer wording will not help. If speech generation causes the wait, cached audio for commonly selected reviewed blocks may help. If the selection step is slow, the candidate set and request path need inspection. A timer around the entire conversation catches the farmer’s experience; timers around each stage explain it.

For Kofi, the important number is simpler: did the answer arrive while he was still standing in the plot with the damaged pod?

Fast answers must still come from reviewed blocks

AgriVoice is designed so the language model selects reviewed content rather than writing new agronomy advice. That boundary matters most when a fast response would sound persuasive.

The system may return a reviewed cocoa disease, pest, or farming-practice block when the question matches it with enough confidence. It may combine approved blocks only where the content design permits it. Pesticide responses retain their mandatory safety content. Three chemical blocks remain held because dosage, re-entry, and pre-harvest details still require Ghanaian agronomist review.

Speed cannot turn those held blocks into advice. It cannot convert unclear speech into a guessed dosage. It cannot make a response safer by sounding fluent in Twi.

This is why latency and safety need to be measured together. A shorter response time is only useful when the selected block is appropriate, or when the system correctly says that a person needs to review the question. The pilot scorecard treats repeated confident wrong answers as a stop or redesign signal, even if replies arrive quickly.

For the questions that need a person, the relevant standard changes. The automated portion should quickly acknowledge the request and route it to the named extension officer. The human response must have an accountable owner and a median response time below one working day. Read more about that handoff in What Happens When a Cocoa Question Needs a Human Answer?.

Reduce delay where it does not change the advice

The safest latency improvements leave the reviewed-answer boundary intact.

AgriVoice can reuse generated speech when the same approved block is selected again, rather than creating identical audio from scratch. It can keep language processing direct for Twi selection if the planned comparison shows it matches or beats the translation route on safety-aware task measures. It can make the acknowledgement and escalation route clear before a human response is ready.

Each proposed improvement needs evidence from real farmer questions, including questions with code-switching, unclear audio, pesticide terms, multiple requests in one note, and questions outside the available content. The team plans to compare direct Twi selection with a Twi-to-translation path on about 40 labelled questions, then repeat the comparison on real speech-recognition transcripts. Exact matches matter, but so do unsafe extra blocks, correct escalation, latency, and marginal cost.

The AIEP Initiative’s work on AI advisory services for smallholder farmers points to a useful discipline: the technical path has to be judged against the advisory job, not treated as a demonstration of model capability. For AgriVoice, a fast answer that selects the wrong reviewed block still fails the farmer.

Keep the conversation open when confidence is missing

Later that morning, Kofi checks his phone again. In this illustrative scenario, AgriVoice has not guessed at a treatment from his short description. It has sent a brief spoken reply that keeps the case open and routes it for follow-up, because the available reviewed blocks do not support a confident answer.

That response is not the same as being left in silence. Kofi knows the question has reached someone, and the extension officer has a defined queue rather than an informal promise that a message will be noticed.

The two-week pilot with 20 to 50 farmers will show whether that distinction holds in practice. Alongside response time, the team will measure successful answers or correct escalations, comprehension, repeat use, safety, escalation operations, and cost per completed question. If latency regularly breaks the exchange, the evidence should point to the actual bottleneck. If an answer is uncertain, the system should preserve the farmer’s question for review, even when waiting is uncomfortable.