In mixed Twi and English speech, “nanso” can reverse the meaning an AI appeared to understand seconds earlier. A conversational system must keep listening beyond the first clear instruction, then interpret what follows “but” before it responds.
At 8:17 p.m. in Kumasi, Ama is standing beside the kitchen sink with a wet teaspoon in one hand and her phone in the other. Her older brother, Yaw, has sent a family voice note about who should collect their auntie the next morning.
“Kwesi can go early. Ɔnim baabi no, na ɔwɔ car. Nanso, ɔrentumi nkɔ bio, enti Ama, please call Kobby.”
The first half sounds settled: Kwesi knows the place, has a car and can go early. Then comes “nanso.” He cannot go after all. Ama needs to call Kobby.
An AI that commits too soon could extract “Kwesi should go” and treat everything after it as a side note. Ama might then confirm the wrong plan in the family group. By the time anyone spots the mistake, their auntie could be waiting with nobody on the way.
For one uncomfortable beat, that is the possible ending.
“Nanso” changes the job of the whole sentence
People do not always place the final instruction first. We explain, soften, reconsider and correct ourselves as we speak.
In Yaw’s note, the details about Kwesi matter because they establish the plan everyone expected. “Nanso” turns that plan around. The words that follow carry the instruction Ama must act on.
This is common conversational work. A sentence can begin with agreement and end with refusal. It can start as reassurance, then reveal a condition. It can sound like permission until the speaker adds the one reason the answer is no.
Code-switching adds another layer. The speaker may set up the situation in English, pivot in Twi, then give the final action in English again. A language selector would interrupt the very pattern the system needs to understand. The language selector Ghanaian speakers don’t need explores why forcing one language per turn can cost more than convenience.
The practical lesson is simple: do not treat the first recognisable clause as the final meaning. Listen for the turn.
A transcript can look right and still miss the point
Suppose the words are transcribed accurately. “Kwesi can go early” appears exactly as spoken. “Nanso” also appears. So does “please call Kobby.”
The transcript can still produce the wrong outcome if the system gives more weight to the opening instruction than the correction that follows. Word recognition alone does not settle who is collecting Auntie.
This is why mixed-language evaluation needs full utterances with real conversational movement. Test a polite lead-in followed by a refusal. Test an English plan corrected in Twi. Test a Twi explanation that lands on an English request. Then check the response, not only the transcript.
Respectful address matters too. A speaker may spend several words acknowledging an elder before disagreeing. Those words are part of the social meaning, while the later turn may carry the decision. A system needs enough context to distinguish courtesy from consent.
Recent Ghana-focused review work has raised concerns about Twi context, respectful address, tonal naturalness and mixed Twi-English speech. Those are connected problems. Conversation depends on how the whole thought develops, including where the speaker changes direction.
The reply should show which instruction survived
With Nkomo, Ama can play or speak the mixed-language message naturally, then interrupt if the reply starts down the wrong path. The useful response is concise and concrete: Kwesi cannot go, so call Kobby.
That reply does more than repeat words. It reveals which part of the message the companion treated as final.
If something goes wrong, the error should appear. A silent failure can leave Ama believing the plan was understood when no reliable answer was produced. Clear errors give her a chance to replay the note, type the key sentence or ask again.
Her privacy choice remains separate from the language task. She can choose whether cloud use is allowed never, each time or for the current session. Her history stays on her device under her control, and turning history off purges it immediately.
These controls matter because family voice notes can contain names, arrangements and details nobody meant to preserve indefinitely. Understanding the sentence should not require vague promises about where the conversation goes.
Test the turn, then test the action
Ama asks one final question: “Enti, hwan na ɔrekɔ?” Who is going, then?
“Kobby.”
She taps back into the family group and calls him before setting the teaspoon beside the sink. The changed state is small and exact: the right person now has the request, and Kwesi is no longer carrying an assignment he already declined.
A good mixed-language test should end the same way. Write down the action implied before “nanso.” Write down the action implied after it. Ask the system one direct follow-up question and see which instruction survives.
Then make the example harder. Let the speaker begin in Twi, switch to English for the plan and return to Twi for the correction. Mixed-language voice-note rehearsal offers a useful companion scenario for testing that flow.
The most revealing question may be only three words long: “Who goes now?”
Comments
No comments yet.