The final English word can prepare a Twi clause, so a natural response depends on following the thought across the switch. A conversational AI has to treat the boundary as part of one sentence, not as a reset button.
In 1999, NASA’s Mars Climate Orbiter approached Mars with a problem that had begun far earlier, in data passed between Lockheed Martin and NASA’s Jet Propulsion Laboratory. One system used pound-force seconds; the other expected newton seconds. The spacecraft was lost because the numbers crossed an interface without their meaning crossing with them.
Arthur G. Stephenson chaired the NASA board that examined the failure. Its report makes the lesson painfully plain: a system can handle each part correctly and still fail at the handoff.
A language switch can carry the same kind of risk, at human scale. The last English word may tell you what kind of Twi thought is coming next. If the listener drops that cue, the reply can sound as though it answered a different conversation.
A switch begins before the first Twi word
Consider: “I will come later because mewɔ adwuma.”
“Because” does important work before the Twi clause begins. It tells the listener to expect a reason. “Mewɔ adwuma” completes that reason: I have work to do.
The meaning is not stored in English on one side and Twi on the other. It unfolds across both. A tool that waits for a clean boundary before it follows the grammar has already arrived late.
The same thing happens in everyday voice messages and chats. Someone may say, “Please tell Ama that, sɛ ɔbɛba a, ɔmfrɛ me.” The English setup points forward to a message. The Twi clause supplies the condition and instruction. Splitting the sentence into separate language tasks can make the reply stiff, incomplete, or strangely certain about the wrong part.
That is why a language picker can create friction at exactly the wrong moment. The thought has moved on while the speaker is still being asked to label it. The Language Picker Nkomo Removed, and What a Pause Can Cost explores what that interruption can cost in a real conversation.
The boundary is where the listener proves attention
A good listener does more than recognize words. They notice what each word makes likely.
After “but,” you expect contrast. After “so,” you expect a consequence. After “if,” you wait for a condition and what follows from it. Those signals still matter when the next clause is in Twi.
This is especially important in speech. A person may pause, correct themselves, speak over an earlier phrasing, then switch languages halfway through the explanation. The natural response is to stay with the developing idea. Nkomo is built for typed or spoken Twi and Ghanaian English, including the ordinary code-switching people use when one language carries the thought better.
The point is not to make a performance of bilingualism. It is to let someone say what they mean without first arranging their sentence for a machine.
A recent ACL 2026 thesis proposal notes that dense within-sentence code-switched speech remains underexplored and that multilingual speech systems can struggle around language-switch boundaries. That boundary deserves attention because it carries grammar, tone, and intent forward.
Listen for the unfinished shape of the sentence
The practical test is simple. Give an AI a sentence whose final English word points into Twi, then see whether its reply respects the relationship between the two clauses.
Try a conditional instruction. Try a reason. Try a correction.
If you say, “Don’t send it yet because…” and finish in Twi, does the response understand that the second clause explains the delay? If you say, “Tell him I agree, but…” and turn to Twi, does it preserve the qualification? If you interrupt while speaking, can you continue the thought without starting again?
A reply can contain correct individual words and still miss the sentence’s direction. That is the language version of the Mars Climate Orbiter problem: the parts arrived, but the meaning did not survive the crossing.
Keep control while you speak naturally
Natural conversation should not require vague promises about what happens to your words. Nkomo gives you explicit cloud-consent choices: never, ask each time, or this session. You can turn off on-device history, which purges it immediately, and use one tap to export your data or delete your account.
Those controls matter when you are practising a difficult message, explaining family news, or recording a thought that moves between Twi and English before you have fully decided how to say it. Before you tap the microphone explains those choices in plain English.
The useful test is small: say the sentence the way you would say it to your sister, colleague, or uncle. Let the final English word lead into the Twi clause. Then listen for whether the answer follows the thought all the way through.
Comments
No comments yet.