A voice assistant that goes quiet when a speaker moves from English into Twi has failed at the moment the conversation becomes natural. A useful assistant must follow the mixed sentence, keep the thread, and respond clearly rather than treating a language change as the end of the request.
In 1970, Apollo 13 had a problem with the same shape, although at a scale nobody would confuse with a stalled voice reply. The command module had square lithium hydroxide canisters. The lunar module used round ones. After the oxygen-tank explosion, the crew needed the lunar module as a lifeboat, but its carbon-dioxide system could not accept the command module’s square canisters. The parts existed. The connection did not.
NASA engineers on the ground worked out an adapter using materials available on the spacecraft. Jim Lovell, Fred Haise, and Jack Swigert used it, and the crew returned safely. Lovell and Jeffrey Kluger document the episode in Lost Moon. The difficult part was not having no solution. It was reaching a boundary where two working systems could no longer work together.
The silence arrives in the middle of a real thought
A person speaking in Ghanaian English and Twi may begin with a practical question in English, add the part that carries the urgency in Twi, then return to English without pausing to announce the switch.
That is ordinary speech. It happens in a family call from Kumasi, a voice note sent from Accra, or a late-night question from a Ghanaian abroad trying to explain a situation back home. The speaker is not changing tasks. They are using the words that fit.
The failure can be painfully small: the assistant begins an answer, the speaker says a Twi phrase to correct or sharpen the question, and then there is silence. Sometimes the system asks for a repeat. Sometimes it returns an unrelated answer. Sometimes it gives no useful indication that anything went wrong.
Silence makes the person do extra work. They start translating themselves. They simplify the question. They cut out the phrase that felt most natural. A conversation that should have helped now asks the speaker to fit the machine’s preferred format.
A language border is a product decision
Natural code-switching is not a mistake to tidy up after the fact. A Ghana-led research team has reported that people commonly mix English and Twi within a sentence, while systems trained on standardised monolingual data can struggle with that natural speech.
That gap shows up in the exact moments people care about. “Can you explain this message, because me nte ase,” carries a clear request. The Twi phrase changes the meaning and the tone of the English around it. Removing it may make the sentence easier for a limited system to process, but it also makes the speaker work around the tool.
The Apollo 13 adapter is useful here because it clarifies the real issue. The square canister was not faulty. The round opening was not faulty. Their mismatch became dangerous because the connection mattered at the worst possible moment.
English and Twi each work. The question is whether the assistant can hold the connection when a person naturally moves between them. For a closer look at what gets lost when an assistant treats switching as an error, read What Happens When an AI Treats English-Twi Switching as an Error?.
The fix begins by naming the failure plainly
Nkomo is built for typed and spoken conversation in natural Twi and Ghanaian English, including the mix people actually use. You can speak with hold-to-talk voice mode, interrupt naturally, hear a reply, or type when that suits the moment better.
That promise needs clear boundaries. Nkomo does not claim that every mixed-language request will be understood perfectly. When something goes wrong, it shows the error instead of swallowing it. That matters because an unexplained silence leaves a person guessing whether their words, their connection, or the assistant caused the break.
The same honesty applies to privacy. A voice conversation can contain family details, health concerns, money worries, or an unfinished thought meant for one trusted listener. Nkomo gives a visible choice over cloud consent: never, ask each time, or this session. History stays on the device under the user’s control, and turning it off purges it immediately. Data export and account deletion are available in one tap.
Those choices answer a different question from language support, but they belong in the same conversation. If a person has to repeat a sensitive question after the first attempt fails, they should know what happens to that second attempt.
Keep the speaker’s thread intact
A good test is simple: say the question as you would say it to someone who knows you. Do not prepare a cleaner version for the assistant. Do not remove the English word that belongs in the Twi sentence. Do not turn a quick spoken follow-up into a formal prompt.
Then watch what happens at the pivot.
Does the assistant continue with the meaning intact? Can you interrupt and clarify without starting the whole request again? If it cannot understand, does it tell you clearly enough to decide what to do next?
Apollo 13’s crew did not need a lecture about square canisters. They needed a connection that worked with what was already there. People speaking Twi and Ghanaian English deserve the same practical respect: a conversation that can carry their words across the point where other assistants fall quiet.
Comments
No comments yet.