A useful Twi-and-English assistant should follow the meaning of a question when the speaker changes languages mid-sentence. It should respond to the whole request, or say clearly when it cannot, rather than making the person begin again in one language.
The question does not arrive in neat language boxes
Afia is standing outside a pharmacy in Kumasi with her aunt’s voice note open and a paper list folded into her palm. The counter is about to close. She holds down the voice button and asks, “Mepa wo kyɛw, this medicine no, can I take it after food, anaa before bed?”
The first response catches “medicine” and “after food,” then answers as if the Twi around them never happened. Afia tries again. This time she changes the order of the sentence, trims out the Twi, and loses the detail she meant to ask about. Her aunt is waiting inside a taxi, and the bad ending is plain: Afia leaves with the wrong understanding, or no answer at all.
That is the small, draining work people do when an assistant expects one language at a time. The question may begin in Twi, reach for an English word because it is the familiar one, then return to Twi for the condition that changes the meaning. The language shift carries part of the thought.
A conversation tool should meet the person where the thought actually happens.
Code-switching often holds the condition that matters
The English noun is rarely the hard part. What can get lost is the “unless,” the “for my child,” the “only if,” or the quiet correction at the end of a sentence.
Think about the difference between “Help me write a message” and “Help me write a message, but don’t make it sound angry.” In mixed Twi and Ghanaian English, that second part may come naturally in another language. If an assistant drops it, it can produce a confident answer to the wrong problem.
That is why a good reply needs to show it has followed the full request. Sometimes the clearest response is brief: restate the condition, then answer. If the request is unclear, ask about the part that is unclear. Do not pretend the missing words never existed.
For voice conversations, this matters even more. People speak while walking, cooking, checking on a child, or standing in a queue. They will interrupt themselves. They will revise a phrase halfway through. They should be able to hold to speak, pause, and continue in the language that comes first.
The practical test is simple: say a question the way you would say it to a cousin. Put the important condition near the end. If the answer ignores that condition, the assistant has not understood the question yet.
For a closer look at why published language labels do not always tell you how a conversation will feel, read [Published language support is not the same as conversational fit]( /blog/published-language-support-is-not-the-same-as-conversational-fit-how-to-assess-a-voice-ai-when-your-everyday-speech-moves-naturally-between-twi-and-ghanaian-english-2fba9f5e/ ).
Keep the conversation moving, without hiding uncertainty
Nkomo is built for natural Twi and Ghanaian English conversation by voice or text. You can ask in the mix of languages you actually use, then hear a reply or read it back. In hands-free voice mode, you hold to speak and can interrupt naturally when you need to correct yourself.
That does not mean every answer should sound certain. A conversational assistant earns trust when it shows errors instead of swallowing them. If something fails, you should see that failure and decide what to do next. An empty pause can sound like understanding when there was none.
Back outside the pharmacy, Afia tries one more time. She phrases the question as she first meant it, including the Twi that marks the timing she cares about. The assistant follows the complete question and gives her a response she can check against the medicine label or take back to the pharmacist. The pressure has changed shape. She has a clear question in hand instead of a half-translated one.
That is a better role for an assistant in a sensitive moment: help someone hold onto their own meaning, while leaving room for the right human decision.
Privacy should be part of the conversation, too
A mixed-language question can become personal quickly. A voice note about symptoms, money, family tension, or immigration paperwork can begin casually and turn private before the speaker notices.
Nkomo makes the cloud choice explicit: never, ask each time, or this session. Its history stays on your device under your control. Turn history off and it is purged immediately. You can also export your data or delete your account in one tap.
Those controls matter because a person should not have to choose between speaking naturally and knowing where their conversation goes. Before you ask the next question, choose the consent setting that fits it. Then speak as you speak.
Comments
No comments yet.