AI should understand a person’s whole thought when they move naturally between Twi and English. Requiring one language per turn makes the person do translation work before the AI can help.
Imagine Adwoa, an invented composite, standing in her kitchen in Kumasi at 6:40 in the morning. The kettle is beginning to whistle, her reading glasses are still on the table, and she is holding down the voice-note button to message her son.
Most of the note is in Twi. Three English words sit inside it: “school portal” and “deadline.”
She does not pause before them. She does not announce that she is switching languages. The words belong to one thought: something on the school portal needs attention before the deadline.
Now imagine an AI splitting that thought at the language boundary. It understands the Twi around the request but mishandles the English words carrying the key action. Adwoa asks it to help her make the message clearer. Instead, it produces a reply about school in general.
Her son could miss the deadline.
For a moment, that outcome remains possible. Adwoa tries again, slower this time, wondering which parts she must translate and whether “portal” even has a replacement that will sound natural in this conversation.
Code-switching carries the meaning
Adwoa’s voice note is not careless speech. It is ordinary communication shaped by family, work, school, radio, church, WhatsApp, and years of moving between languages without asking permission.
The English words are not foreign objects dropped into a Twi sentence. They may carry the most precise meaning available for the moment. “School portal” points to a specific place and task. “Deadline” supplies urgency. The Twi around those words carries relationship, tone, context, and what Adwoa expects her son to do.
Understanding the sentence requires understanding how those parts work together.
This is why linguistic purity is a poor design goal for conversational AI. A clean, single-language input may be easier for a system to process, but the burden shifts to the person speaking. Before asking for help, she must edit herself. She must choose one language, translate the awkward parts, and hope the new sentence still sounds like her.
That extra work changes who benefits from the technology. The most comfortable experience goes to the person willing and able to speak in the system’s preferred format.
A conversation should follow the speaker
With the deadline still hanging over her, Adwoa opens Nkomo and holds to speak. She says the thought as it came the first time, Twi flowing around “school portal” and “deadline.” She hears a concise response and interrupts when she wants to adjust the wording.
The important turn is simple: she can continue from the meaning she already has, rather than rebuilding the sentence for the machine.
Nkomo supports natural Twi and Ghanaian English conversation through voice or text. It is made for the mixed language people actually use, including moments when a sentence begins in one language, borrows the exact word it needs from another, and returns without explanation.
GhanaNLP’s Twi speech-synthesis model offers further evidence that this requirement belongs at the centre of product design. The model was designed to read English words inside Twi text. That choice recognises a practical truth: code-switching is part of the input, not noise to remove before processing begins.
No conversational system should claim perfect understanding. When something goes wrong, the error should be visible. Nkomo does not swallow failures and leave the speaker guessing whether the message was heard, misunderstood, or ignored.
Privacy should not require translation either
A natural conversation can contain personal details before the speaker notices how much she has shared. A family voice note may mention a child, a payment, a health concern, or a disagreement that belongs inside the family.
Clear privacy controls matter because people should not need technical language to decide what happens next.
Nkomo makes the cloud choice explicit: never, ask each time, or allow it for the current session. Conversation history stays on the device under the person’s control. Turning history off purges it immediately. Data export and account deletion take one tap, and sign-in uses a magic link, so there is no password to remember.
These controls support the same principle as code-switching support. The product should adapt to the person’s decision. The person should not have to learn the product’s hidden assumptions.
Keep the whole thought intact
Later that morning, Adwoa’s son receives a clear message with the school portal and deadline preserved. The kettle has gone quiet. Her glasses are back in their case. She has moved on to the next part of her day without turning her own sentence into a translation exercise.
That is the practical test for conversational AI in Ghana and across the diaspora. Give it the sentence someone would actually say to a sibling in Accra, a mother in Kumasi, or a cousin abroad. Keep the pauses, borrowed words, and mid-sentence switches.
Then ask one question: did the system understand the whole thought, or did the speaker have to become someone else first?
Comments
No comments yet.