A helpful voice assistant should keep the unfinished meaning of a mixed Twi and English message when a person returns after an interruption. The language change is often part of the thought, so the reply has to follow the intention that was already in motion.
The thought was still waiting by the phone
At 6:18 p.m., Auntie Adwoa has one hand on a bowl of sliced kontomire and the other around her phone. She is recording a message for her nephew, who is due to travel with a parcel for his cousin. “Kɔfa the blue bag no, na the documents inside no, please make sure…”
The doorbell rings.
She puts the phone down to answer it. A neighbour has come to return a dish, and the conversation takes a few minutes. When Auntie Adwoa comes back, she does not restart from the beginning. She picks up the thread where her mind still is: “...and tell him the envelope should stay with you, because he may leave before I get there.”
The risk is clear. If the assistant treats that English sentence as a new instruction, it may miss that the envelope belongs with the documents in the blue bag. Her nephew could leave with the parcel but without the one item Auntie meant to protect.
That is a small household interruption with a real consequence. The task is not to admire the switch from Twi to English. The task is to remember what “the envelope” means in the message she had already started.
An interruption should pause a conversation, not erase it
People rarely speak in tidy turns at home. Someone calls from the kitchen. A child needs help finding a shoe. A delivery arrives. The phone is put down, then picked up again, and the sentence resumes in whichever language feels quickest.
Auntie Adwoa might return with “Enti don’t give it to anybody else until he calls.” That “enti” carries the earlier point forward. It tells you this is a consequence, not a fresh topic. A useful assistant needs to hold the thread long enough to understand it.
This matters most when the message contains instructions, dates, names, addresses, or conditions. The danger is not only a wrong translation. A system can hear every word and still lose the relationship between them.
In the composite scene above, the key instruction has three connected parts:
- the blue bag contains the documents;
- the envelope belongs with that set of items;
- the nephew must hold it until a call confirms the handover.
If the assistant drops the first part after the doorbell interruption, it can produce a clean-looking summary that quietly changes the meaning. That is how a message becomes harder to trust.
Nkomo is built for natural Twi and Ghanaian English conversation by voice or text, including the mixes people already use. You can hold to speak, interrupt naturally, then continue the thought. The aim is to let the conversation sound like a conversation, while keeping the context needed for the reply.
Code-switching can carry the instruction itself
A person may start in Twi because the relationship, setting, or first part of the message sits naturally there. They may switch to English for a name, a travel detail, an instruction they want to make crisp, or simply because that is how the sentence comes out.
Flattening that mix can remove useful signals. The English clause after the interruption may be the exact condition that gives the Twi opening its force. The earlier Twi phrase may tell us what “it,” “there,” or “that one” refers to.
The 2026 GhanaNLP Community dataset describes natural conversational Twi audio as containing the Twi and English code-switching heard in Ghanaian speech. Its machine-generated transcripts are not human-verified ground truth, which is an important caution. A transcript can look complete while still getting a word, name, or reference wrong.
That is why a good workflow leaves room to check the original spoken message. If something is unclear, the assistant should surface the problem instead of silently filling in a confident guess. Nkomo shows errors rather than swallowing them, so a person can decide what to repeat, correct, or send another way.
For a related example of what can disappear when mixed speech is rewritten too neatly, read The English-to-Twi Switch in Your Voice Note, and What Rewriting It Can Flatten.
Give people control over the conversation they are continuing
Continuity matters, and so does control over what is retained to provide it. A household message may contain a phone number, a family arrangement, a health detail, or a travel plan. People deserve a clear choice about what reaches the cloud and what stays on their device.
With Nkomo, cloud consent can be set to never, ask each time, or this session. Conversation history stays on the device under the person’s control; switching history off purges it immediately. Data can also be exported or the account deleted in one tap.
Back in Auntie Adwoa’s kitchen, the bowl of kontomire is still on the counter when she listens back. This time, the message keeps the blue bag, the documents, and the envelope together. She hears the instruction in the order she meant it, then sends it before dinner gets cold.
Comments
No comments yet.