What Happens When a Farmer Wants Their Voice Recording Deleted?

Responsible speech collection gives farmers a clear choice, records what they agreed to, and lets them ask for their recordings to be deleted. Better language models need real field voices, but that need never gives a project ownership of a person’s voice.

In 1951, Henrietta Lacks was treated for cancer at Johns Hopkins Hospital in Baltimore. Cells taken during her treatment became the HeLa cell line, widely used in medical research for decades. Her family later learned how extensively those cells had been used. The history is documented by the National Institutes of Health and Rebecca Skloot’s The Immortal Life of Henrietta Lacks.

The lesson travels well because the mechanism is familiar: material collected in one vulnerable moment can acquire new value elsewhere, long after the person who provided it has lost sight of it. A farmer’s voice note about a diseased cocoa pod is not a cell sample. It still carries identity, language, location clues, and the right to decide what happens next.

Consent belongs in the conversation

A farmer who sends a voice question should hear a plain explanation before that recording can be retained for evaluation or improvement. The explanation needs to say what will happen to the audio, whether it will be transcribed, who can access it, how long it may be kept, and how to withdraw permission.

“Help improve the service” is too vague. A person needs a concrete choice: use this recording to answer my question only, or allow it to be retained for carefully limited review and model evaluation.

That distinction matters in Asante Twi. AgriVoice needs real farmer speech because read speech does not capture the pauses, code-switching, background noise, local vocabulary, and hurried phrasing that appear in a WhatsApp voice note. But a useful training set cannot begin with an assumption that every question is available forever.

Consent should be voluntary, separate from receiving help, and easy to decline. A farmer must still be able to ask about a suspected cocoa disease, hear a reviewed answer, or reach an extension officer without agreeing to contribute their audio for future work.

Deletion must work after the useful moment has passed

The difficult test comes later. A farmer may consent when they need an answer quickly, then reconsider after discussing it with family, a cooperative leader, or an extension officer. Responsible collection gives that person a simple way to request deletion without explaining why.

Deletion must cover more than the original audio file. Teams should map where the recording went: the message record, transcript, evaluation set, reviewer queue, backups, and any derived training material. Some records may need to remain in a restricted form for safety or operational accountability. That boundary should be explained before consent, rather than discovered after a request arrives.

For a small pilot, this can be practical. Assign each consented recording a clear identifier. Keep the consent decision beside it. Test the deletion path before farmer exposure. Confirm that staff can find the recording, remove it from active review and evaluation stores, and record that the request was completed without keeping unnecessary personal detail.

The same discipline helps the product. If a team cannot explain where one voice note lives, it cannot credibly promise that the note can be removed.

Collect only what the next decision requires

A better model does not require every detail surrounding a farmer’s question. Start with the minimum needed to measure whether speech recognition and safe answer selection work: the audio, an agreed transcript, consent status, and a broad category such as disease, pruning, or pesticide safety.

Avoid turning a voice note into a hidden profile. Phone numbers, precise locations, household details, and farm identifiers create more risk and often add little to an early speech evaluation. When a question needs a human response, the [named extension officer]( /blog/the-named-extension-officer-agrivoice-needs-and-what-an-unanswered-question-could-cost-aa6bb473/) may need contact details to follow up. The model evaluation process usually does not.

The strongest collection plan separates these purposes. A service record helps answer or escalate the farmer’s question. A consented, de-identified evaluation record helps measure whether the system understands field speech. The two should not become one large, permanent database by accident.

Trust is part of the field evaluation

AgriVoice’s pilot should test consent and deletion behavior alongside speech recognition, comprehension, safety, response time, and repeat use. A model that performs well only because people felt unable to refuse has failed a more basic test.

Henrietta Lacks’s story remains difficult because scientific value grew from material whose owner had no meaningful control over its future use. Field voices can help build better Ghanaian-language services. The durable way to earn them is to make every contributor’s choice visible, limited, and reversible.

Before collecting the first pilot recording, run one deletion request from start to finish. Treat that result as seriously as an answer-rate metric.