One-language-per-turn design makes the speaker do extra work before an AI can help. A Twi and Ghanaian English companion should understand the mix people already use, without asking them to translate themselves into a settings menu.
At 7:18 p.m. in a kitchen in Kumasi, Auntie Esi has one hand on a bowl of peeled kontomire and the other around her phone. Her nephew is leaving early the next morning, and a parcel still needs to reach him before he goes.
She holds the button and starts quickly: “Please tell Kofi sɛ ɔnnkɔ ntɛm, because the parcel no, me de maame no nkyɛn already, but he has to call before he comes. Na the address no, I sent it yesterday.”
Then she stops.
A tool that expects one language at a time creates a new job at the worst moment. Should she say the whole thing again in English? Should she force the Twi parts into words that feel less precise? Should she split one thought into two messages and hope the important condition survives?
If Kofi gets only “the parcel is ready,” he may leave without calling first. The parcel could stay behind after his trip begins. The pressure sits in the missing detail, right where Auntie Esi switched languages.
The switch often carries the point
Code-switching is not a mistake waiting to be cleaned up. It can show urgency, soften an instruction, add emphasis, or reach for the phrase that lands first.
In Auntie Esi’s voice note, “ɔnnkɔ ntɛm” carries more than “don’t go early.” It has the directness of a family instruction, delivered in the language that makes the relationship clear. “The parcel no” points back to the exact thing everyone has already been discussing. The English words help carry the logistics. The Twi words carry the pressure and the family context.
When an assistant accepts only one language per turn, it asks the speaker to flatten that structure before the conversation can begin. That can make a person slower and less certain, especially when they are trying to remember an address, explain a condition, or get wording right before a call.
The same problem appears in typed messages. A person may begin in English, switch to Twi for the part they want to stress, then return to English for a name or next step. Treating that as a formatting error misses the thought behind it.
For a closer look at what rewriting can lose, see The English-to-Twi Switch in Your Voice Note, and What Rewriting It Can Flatten.
Speech tools should meet people where they speak
Auntie Esi does not need a lesson in language purity. She needs to say what she means while the pot is on the stove and the person with the parcel is still reachable.
Nkomo is built for natural Twi and Ghanaian English conversation by voice or text. You can hold to speak, hear a reply, and interrupt naturally when you remember the detail you left out. That matters because real speech rarely arrives as a polished paragraph. It comes with corrections: “Wait, add that he should call first.” It comes with a switch halfway through a sentence because that is how the thought arrived.
The useful test is simple: say the sentence you would say to your sister, colleague, or uncle. Do not first convert it into the version you think a machine prefers. If the tool needs you to perform a different voice before it can understand, the tool has moved the work onto you.
Language technology is beginning to take Ghanaian languages more seriously. GhanaFact has reported plans involving Google, the University of Ghana, and partners to develop speech-recognition and AI tools for 12 Ghanaian languages, including Asante Twi and Akuapem Twi. That work matters. The everyday test still happens in small moments, when one person needs to speak naturally and be understood.
Privacy choices should be plain too
The same respect applies to data. Someone speaking a mixed-language message may be sharing family plans, health worries, money questions, or a private disagreement. They deserve a clear choice about what reaches the cloud and what remains on their device.
Nkomo makes that choice explicit: never, ask each time, or this session. Its history stays on the device under the person’s control, and turning history off purges it immediately. Data export and account deletion are available in one tap. If something goes wrong, the app shows the error rather than pretending the message was handled.
Those details belong close to the conversation because they affect whether a person feels comfortable speaking freely. Privacy language should be as understandable as the reply itself.
Let the person keep the natural version
Auntie Esi tries again, this time without rehearsing. She says the mixed message as it came to her, then adds the one detail she nearly lost: Kofi must call before he comes.
The next morning, the parcel goes with him.
That is the standard worth keeping: the person should leave a conversation with their meaning intact, their choices clear, and no password or hidden failure standing between them and the next thing they need to say.
Comments
No comments yet.