Translation tool vs. conversation partner: why Khaya AI proves Twi speech tech works, and why that's a different problem from holding an ongoing, code-switched back-and-forth

Khaya AI shows that Twi speech, translation and language switching can be supported in an AI tool. Holding a useful ongoing conversation asks for a broader set of abilities: remembering context, following a mixed sentence, knowing when you have changed your mind, and responding in the language mix you used.

Start with the job you need done

Use a translation tool when you have a clear source message and want help moving it into another language. You might dictate a Twi sentence, translate it, check the wording, then send the result. The task has a defined beginning and end.

Use a conversation partner when the message is still taking shape. Perhaps you begin, “Mepɛ sɛ me reply this landlord message, but I don’t want it to sound rude.” Your next turn may add the rent amount, change the tone, or ask for a shorter version. The useful answer depends on the earlier turns as much as the latest sentence.

That distinction matters because a tool can handle Twi speech and still leave you doing the work of rebuilding context every time you speak. Khaya AI’s support for Twi speech, translation and language switching is meaningful progress. It gives people a reason to expect more from language technology than English-only input. It also sets up a separate question: can the tool stay with the way you actually talk?

Test the second and third turn, not only the first

A first-turn test is easy to pass. Say one sentence in Twi, English, or both, then see if the response is understandable. That tells you something, but it does not tell you how the tool behaves in a real back-and-forth.

Try a three-turn test instead:

1. Say: “Help me write a short message to my auntie. Meka sɛ I’ll come on Saturday.”
2. Follow with: “Make it warmer, but don’t mention the time again.”
3. Then switch naturally: “Add that I’ll bring the documents too, na mma no nyinaa nkɔ so.”

A conversation partner should keep the original purpose, apply the change in tone, avoid reintroducing details you removed, and understand that the last phrase belongs inside the same request. If you must repeat the entire message after every adjustment, you are managing the conversation instead of using it.

For a closer look at what can go wrong in a mixed exchange, read What Happens When Your Voice Assistant Loses the Thread Between Twi and English?.

Code-switching carries meaning beyond vocabulary

People do not always switch between Twi and Ghanaian English because they lack a word. A switch can make a message softer, more direct, funnier, more respectful, or more precise.

Take: “Please, tell him sɛ mepɛ the receipt today.” The English word “receipt” may be the practical word the speaker wants. Forcing a replacement can make the message sound unnatural or create uncertainty. The important task is to follow the intended meaning and register without making the speaker choose one language for the whole turn.

The same applies to corrections. “No, not today. Friday morning na ɛyɛ better” is not a fresh translation prompt. It changes a detail from the previous turn. A useful response needs to retain the request, replace the date, and keep the speaker’s tone.

This is why language switching is a foundation, not the full destination. Translation and speech support make individual turns possible. Conversation depends on continuity across turns.

Voice changes the expectations

Speech adds another layer. In a typed chat, you can reread and edit before sending. In voice, people pause, restart, add English halfway through, and interrupt when the answer is heading in the wrong direction.

A voice tool should make those repairs feel possible. Nkomo’s hands-free voice mode lets you hold to speak, interrupt naturally, and hear the reply. That matters for a request such as, “Explain it simply... wait, make it for my mum, in Twi mostly.” The correction is part of the conversation.

Keep answers short when you test voice. Long replies make it harder to notice whether the system followed your correction, and they give you fewer chances to steer the exchange. Ask for a two-sentence draft, a shorter rewrite, then one change in language or tone.

Check what happens to a sensitive conversation

A language tool can be useful and still require a privacy decision. Voice messages may include family matters, finances, health concerns, or work details. Before using any AI for those conversations, find out what reaches the cloud, what remains on the device, and what control you have afterward.

Nkomo makes that choice explicit: you can set cloud consent to never, ask each time, or this session. Its history is on-device, and turning history off purges it immediately. You can also export your data or delete your account in one tap. Nkomo does not claim offline conversation, so cloud consent is especially worth checking before you share a sensitive voice note.

The practical next step is simple: run the same three-turn, mixed-language test on the tools you are considering. Include one correction, one change of tone, and one detail you do not want repeated. Then choose the tool that keeps the thread, speaks in a way that sounds like you, and gives you privacy controls you can understand before you talk.