An assistant can understand every English sentence and still give the wrong answer when the speaker’s real purpose arrives in Twi. For Twi-English speakers, meaning often lives across the whole conversation, especially in the final turn that changes what the earlier details were for.
In 1999, NASA lost the Mars Climate Orbiter as it approached Mars. The Mars Climate Orbiter Mishap Investigation Board found that one team’s ground software used pound-seconds while NASA’s navigation software expected newton-seconds. Each side had produced data. The data did not mean the same thing inside the shared system, and the spacecraft was lost.
The final switch can change the whole request
Imagine someone typing in English: “I sent the money yesterday. The receipt has a different name. My auntie says she has not seen it.”
An assistant could correctly explain what a receipt is, suggest checking the transfer reference, and draft a polite message. Then the person adds: “Mepa wo kyɛw, mma menka nkyerɛ no sɛ wanya sika no. Me pɛ sɛ mehu sɛ ɛbɛkɔ hɔ ansa na meatu balance no.”
The decisive point is no longer a general question about receipts. They need to verify that the money will arrive before they pay the remaining balance. A reply that continues explaining receipt names has processed the sentences while missing the job.
That is why language support has to cover more than translating a final phrase. The assistant has to carry the earlier English details into the Twi turn, recognise the shift in urgency, and answer the actual next step.
A correct translation can still miss the point
Code-switching often does real conversational work. A speaker may use English for the transaction details, then switch to Twi for what they are worried about, what they do not want said aloud, or what has become urgent.
“Let me know if it has gone through” can be a routine request. “Mma menka nkyerɛ no sɛ wanya sika no” can reveal a different constraint: do not tell her that the money has arrived yet. Context matters because the requested wording, the timing, and the person involved all change the useful answer.
This is the same failure shape NASA documented in its Mars Climate Orbiter report. A value can look valid inside one part of a process and still cause harm when its meaning does not carry into the next part. In conversation, a word-for-word interpretation can look polished while sending someone toward the wrong action.
A useful assistant should be able to say, in plain language, “You want to confirm the transfer is on its way before you pay the balance, and you do not want the recipient told it has arrived until you have proof.” That response shows it has held onto the whole thread.
Conversation needs room for correction
Real speech changes direction. Someone may start with a message they want to send, interrupt halfway through, and add the detail that matters most. A voice assistant that treats every turn as separate makes that correction harder.
Nkomo supports natural Twi and Ghanaian English in text or voice, including the mix people use in ordinary conversation. In hands-free voice mode, you can hold to speak, interrupt naturally, and hear the reply. That makes space for the practical correction: “Wait, the issue is the balance,” or “No, don’t mention the receipt name.”
The goal is not to promise perfect understanding. It is to make the assistant’s interpretation visible enough that you can correct it before acting on a wrong answer. If something fails, Nkomo shows the error instead of silently swallowing it.
For a related example of how one phrase can change a family payment question, see Twi-English Conversation: How Afua Clarified a Family Payment Question.
Privacy belongs in the same conversation
Payment questions, family messages, and voice notes can carry details people do not want retained by default. The person asking for help should be able to decide what reaches the cloud and what remains on their device.
Nkomo offers explicit cloud-consent choices: never, ask each time, or this session. Its on-device history can be turned off, and turning it off purges that history immediately. You can also export your data or delete your account in one tap.
Those controls do not solve a misunderstanding by themselves. They give you a clearer boundary around a conversation that may contain sensitive context. When the key meaning arrives late, in Twi, and with a detail you would rather keep close, that boundary matters.
The Mars Climate Orbiter was lost because a system failed to preserve meaning across a handoff. A Twi-English conversation deserves better handling at its own handoffs: from English to Twi, from voice to reply, and from a first request to the actual decision that follows.
Comments
No comments yet.