Before you accept payment for a voice-data call, ask where the recording will go, who can access it, how long it will be kept, and whether it can be deleted. A paid invitation to speak Twi for an hour may be legitimate work, but payment does not answer the privacy questions for you.
Efia had the invitation open on her phone at 8:40 p.m., sitting at her kitchen table in Kumasi with a half-finished bowl of waakye beside her. The post promised payment for completed conversations and asked for fluent Twi speakers. She could speak easily, switching into Ghanaian English whenever a phrase came faster that way. The money would help with a bill due soon.
Then she pictured the call itself. An hour can hold plenty: her mother’s health worry, a cousin’s awkward marriage plans, the name of a shop she uses, a voice that relatives would recognise in two seconds. She could finish the recording, receive payment, and still have no clear idea where that voice might travel after she hung up. The uncomfortable possibility sat there: a permanent copy of a personal conversation, reused beyond the purpose she thought she had agreed to.
Efia stopped before tapping “apply.”
Payment tells you the price, not the destination
A recruitment post can make the exchange sound simple: speak naturally, complete the call, get paid. For language-data work, the actual exchange can be much bigger. Your voice, words, speech patterns, and any personal details that enter the conversation may become data for a process you cannot see.
Start with the plain question: where will this recording go?
Ask whether the call is stored, transcribed, shared with another company, reviewed by people, or used to train or evaluate a system. Ask whether the organisation removes names, phone numbers, addresses, and other identifying details before further use. Ask how long it keeps the raw audio and whether you can withdraw consent later.
These questions matter especially when the invitation asks for natural speech. Natural Twi is rich with context. A casual “Medaase, but let me explain” can carry family history, work pressure, location clues, and relationships without sounding private in the moment. The recording can preserve more than the task description suggests.
A clear answer may still lead you to take the work. That is a valid choice. The point is to make it with your eyes open.
Read consent language as if the recording were yours to protect
Consent language often appears near the bottom of a form, surrounded by checkboxes and terms that feel designed for speed. Slow down there.
Look for the exact purpose of the recording. “Language research” is broad. “Used to assess speech recognition for a defined project” gives you more to work with. If the terms allow future uses, transfer to unnamed partners, indefinite storage, or broad commercial rights, recognise what that means before agreeing.
It also helps to separate three different questions:
- Can the organiser collect this recording for the call?
- Can it keep a copy after the payment is made?
- Can it use or share that copy for another purpose later?
One “yes” should not blur into all three.
Efia wrote down four questions before replying. She did not need legal language. She asked where the audio would be stored, who would hear it, how long it would remain available, and what would happen if she later asked for deletion. If a recruiter cannot answer plainly, that uncertainty is useful information. You do not have to fill the silence with trust.
Protect the conversation before the microphone turns on
Even with satisfactory answers, treat a paid recording like a public-facing task with a private boundary. Prepare a version of natural speech that does not expose details you would regret hearing played back elsewhere.
Use fictional names when the task permits. Avoid account numbers, addresses, workplace details, school names, medical information, and stories about other people who did not consent to appear in the recording. If the prompt requires personal examples, keep them general. “A family member was unwell” carries the point without naming the person, clinic, or condition.
That boundary is not a failure of fluency. It is good judgment.
For everyday AI conversations, the same habit applies. Before speaking, know what reaches the cloud and what stays on your device. Nkomo gives you explicit cloud-consent choices: never, ask each time, or this session. You can also turn off on-device history, which purges it immediately, and export or delete your account in one tap. Those controls support a simple decision at the moment it matters, rather than asking you to remember a privacy setting from months ago.
For a fuller look at the difference between cloud consent, local history, export, and deletion, read [Privacy is not one feature]( /blog/privacy-is-not-one-feature-a-plain-english-comparison-of-offline-ai-claims-cloud-processing-consent-local-history-export-and-deletion-for-african-language-voice-tools-10c89427/ ).
A good invitation leaves room for a real choice
The next evening, Efia received a reply. It explained the recording process in plain language and answered her questions directly. She read it twice, then decided what she was comfortable contributing. During the call, she spoke warmly and naturally, but kept her examples clear of the family details that had made her pause the night before.
Her waakye bowl was gone by then, replaced by a glass of water and a short list beside her phone. The list had done its job. She had not guessed what would happen to her voice. She had asked.
Before any voice-data call, make your own short list. Keep it close enough to use before the microphone comes on.
Comments
No comments yet.