A speaker should be able to move between Twi and English without stopping to choose a language setting. The AI should carry the burden of following the conversation, because code-switching is part of how people naturally speak.
Imagine Auntie holding down the record button:
“Mepa wo kyɛw, call me when you reach. Na the papers no, fa bra, because they’ll ask for them.”
She does not pause before “call me.” She does not announce that Twi has ended and English has begun. The sentence moves where the meaning needs it to move.
A language selector would ask her to reorganize that thought for the machine. Choose Twi, and the English may become the problem. Choose English, and the Twi may be treated as noise. Either way, Auntie ends up doing translation work before she can ask a simple favour.
The cost of making people match the system
In 1999, NASA’s Mars Climate Orbiter approached Mars after a journey of more than nine months. The spacecraft was meant to enter orbit, but communication was lost as it passed behind the planet. It never reappeared.
The investigation found a mismatch at the boundary between two systems. Lockheed Martin software produced thruster data using pound-seconds. NASA’s navigation software expected newton-seconds. One side supplied imperial units; the other interpreted the figures as metric. The spacecraft was lost.
NASA’s Mars Climate Orbiter Mishap Investigation Board Phase I Report documented the mismatch and the process failures around it. The lesson reaches beyond spacecraft: when two valid ways of expressing information meet, the interface must handle that boundary clearly. Quietly assuming that everything arrives in the expected format creates risk.
Human conversation is less rigid than a navigation calculation, but the design lesson still applies. When someone naturally mixes Twi and English, forcing them to preselect one language moves the system’s coordination problem onto the speaker.
That burden appears small on a screen. In conversation, it changes behaviour. The speaker may shorten the message, replace the word that came first, repeat a sentence, or abandon voice and type instead. Each adjustment helps the machine while making the exchange less natural for the person.
Code-switching carries meaning
Twi and English do not always sit in separate boxes in Ghanaian speech. A conversation can begin in English, shift into Twi for emphasis, return to English for a work term, then end with a familiar Twi expression.
Those choices can carry tone and relationship. Twi may make a request feel warmer. English may supply the phrase used at work or school. A mixed sentence may simply be the quickest and clearest way to say what needs saying.
This is why the language selector can become more than an extra tap. It asks the speaker to predict which language will dominate a thought that has not yet been spoken. It also suggests that switching languages is an exception requiring correction.
Nkomo is designed for typed or spoken conversation in natural Twi, Ghanaian English, and the mix people actually use. In voice mode, you hold to speak, hear the reply, and can interrupt naturally. The aim is direct: let the conversation move without asking the speaker to police every language boundary.
The same principle matters during evaluation. A system should be tested with mixed-language speech, Ghanaian vocabulary, names and ordinary sentence patterns, rather than only clean, single-language prompts. The current research context reflects this need by pairing Twi and English across monologue, narrative, dialogue and story-oriented styles grounded in Ghanaian news topics, vocabulary and named entities.
For a closer look at the practical stakes, Conversational AI for Ghana: Why Twi and English Must Work in the Same Sentence explores why mixed-language understanding belongs at the centre of the experience. The Language Selector Ghanaian Speakers Don’t Need, and What It Can Cost examines the friction created when software asks people to choose too early.
Control should remain explicit where it matters
Removing a language choice should not mean hiding every choice. Some decisions belong with the system. Others belong clearly with the person.
Nkomo makes cloud consent explicit: never, ask each time, or allow it for the current session. Conversation history stays on the device under the user’s control. Turn history off and it is purged immediately. Data export and account deletion each take one tap. When something goes wrong, the error appears instead of disappearing silently.
This division matters. The AI can follow the language without demanding constant configuration. The person still decides what reaches the cloud, what remains on the device, and how long local history stays there.
Good defaults reduce needless effort. Clear controls protect meaningful choice.
Design for the sentence people actually say
Before adding a language selector, test the sentence that crosses its boundary. Ask a Twi and English speaker to send a voice note as they normally would. Keep the pauses, borrowed terms, changes of direction and mid-sentence switches. Then watch where the product asks the person to accommodate the interface.
The Mars Climate Orbiter investigation showed what can happen when a boundary depends on everyone silently using the same convention. Conversational software has a simpler remedy: accept the speaker’s natural input, show failures clearly, and reserve deliberate choices for privacy and control.
Auntie already knows how to say what she means. The language setting should not become her homework.
Comments
No comments yet.