An AI should let a person ask in Twi, add the English word that fits naturally, and continue the same thought. Restarting a sentence because a tool rejects an ordinary language switch turns a simple question into work.
In 1945, the Nuremberg trials faced a version of that problem with far higher stakes. Proceedings at the Palace of Justice in Nuremberg had to be understood across English, German, French, and Russian. If every speaker had waited for one language to finish before another began, the trial would have moved at an unworkable pace.
Léon Dostert, who helped organise the interpreting system, worked with a team using simultaneous interpretation so people could follow the proceedings as they happened. The approach was still demanding. Accuracy mattered, timing mattered, and a missed phrase could change what listeners understood. The United States Holocaust Memorial Museum documents how interpretation became central to making the multilingual trial function.
Auntie asking a question at home has different stakes, but the principle holds. Language should carry the thought forward. The tool should not make her choose between speaking naturally and being understood.
The English word is often the right word
A conversation in Ghana can move from Twi to English inside one sentence because that is how the point comes out clearly. Someone might ask, “Sɛ me pɛ sɛ me change appointment no a, dɛn na menyɛ?” The word “appointment” may be the word she uses every time. Replacing it just to satisfy an assistant makes the sentence less natural, not more correct.
That small interruption carries a cost. She may simplify the question, leave out context, or decide the tool is only useful when she has time to translate herself first. For a question about a family message, a recipe, a school form, or a health appointment, that extra effort can make the assistant feel distant.
Nkomo is designed for typed and spoken conversations in natural Twi and Ghanaian English, including the mix people actually use. A person can hold to speak, interrupt naturally, and hear a reply without rebuilding the sentence around the assistant’s limits.
A restart breaks more than the sentence
When a voice assistant falls silent at an English-Twi switch, the immediate problem is obvious. The person has no answer. The quieter problem is confidence. Next time, they may avoid speaking Twi, avoid voice mode, or stop asking the question altogether.
That is why clear behaviour matters when a conversation goes wrong. Nkomo shows errors rather than swallowing them. A visible error gives the person information: the request did not go through, and they can decide what to do next. Silence leaves them guessing whether the tool heard them, sent something, or forgot the question halfway through.
The same care belongs in privacy choices. A person asking about a private family matter should be able to see whether cloud use is set to never, ask each time, or this session. They should also control local history, turn it off, and have it purged immediately. Privacy is part of the conversation, especially when the conversation begins in the language people use with family.
For a closer look at the cost of a voice tool failing at a switch, read What Happens When a Voice Assistant Falls Silent at an English-Twi Switch?.
Natural speech keeps the useful detail intact
The best question is usually the one someone says first. It contains the names, the urgency, the word they know from work or the clinic, and the Twi phrase that gives the situation its actual meaning.
Consider the difference between asking, “How do I change an appointment?” and explaining that a family member has already travelled, the time has moved, and the appointment must be changed before a particular conversation. The second version has the detail needed for a useful reply. If the speaker gets stopped at “appointment,” that detail may never arrive.
This is where language support becomes access to information. People should be able to bring their whole question, including ordinary code-switching, rather than compress it into the language pattern a tool happens to accept.
The Nuremberg interpreters had to preserve meaning while speech moved across languages in real time. A voice-and-text companion has a smaller job, yet the human expectation is similar: hold on to the thread. Let the person finish. Respond to what they meant.
Make the first attempt count
For everyday use, the practical test is simple. Ask the question as you would ask your sister, colleague, or auntie. Use Twi where Twi is natural. Use English where English is the word you normally reach for. Speak it or type it in one go.
Then pay attention to what happens with your information. Choose the cloud-consent setting that fits the question. Keep history on only when you want it available on the device. Export or delete your data in one tap when you need to.
A good conversation tool earns its place by removing a small, repeated frustration. Nobody should have to start again because the sentence sounded like home.
Comments
No comments yet.