Akosua’s medicine detail was blurred. Her cousin could pick the wrong one.

Forcing a Ghanaian voice note into one language label can change its meaning before anyone responds. Mixed Twi and Ghanaian English carry meaning in the switch itself, and a transcript that guesses the wrong frame can turn a clear request into a confusing one.

At 6:40 one evening in Adum, Akosua stood outside a pharmacy with her phone pressed to one ear and a folded shopping list in the other hand. Her mother was waiting at home, and Akosua needed her cousin to collect a medicine before heading across town.

She sent a voice note: “Please, me maame no medicine no, pick it up before you pass through. The red one, na ɔkaa sɛ ɛno na ɛboa no paa.”

The message moved naturally between English and Twi. The English words set the task. The Twi carried the detail about which medicine her mother had said helped most.

A transcription tool labelled the whole note as English. The result was a string of almost-familiar words with the important detail blurred. Her cousin replied, “Which red one?”

The wrong medicine was now a real possibility. So was a wasted trip, an anxious mother at home, and a second search after the shop had shut for the day.

A language label can decide what the system hears

Language detection often starts with a forced choice: English, Twi, or another language. That choice sounds tidy. Everyday speech rarely is.

A person may begin in Ghanaian English, switch to Twi for a phrase that feels more exact, then return to English for a name, amount, or instruction. The switch may signal intimacy, urgency, correction, or a detail that does not land the same way in translation.

When a system treats that voice note as one language from start to finish, it can spend the rest of the interaction trying to make the words fit its first guess. The output may look polished enough to trust. That is part of the danger.

Akosua’s cousin did not need a perfect linguistic analysis. He needed to know which medicine to pick up. A transcript that turns a useful detail into uncertainty has already failed the conversation.

Code-switching has a job to do

People mix languages for practical reasons. One word comes faster in Twi. A family instruction sounds warmer in English. A phrase carries the right weight only in the language where it is usually said.

Treating every switch as an error asks people to edit themselves before they can be understood. It also adds work at the worst moment: when someone is walking, cooking, travelling, caring for a relative, or trying to send one clear instruction before a battery runs down.

That is why mixed-language input needs to be handled as ordinary speech, not as a problem to clean up. The goal is not to make a person sound like they stayed inside one language. The goal is to preserve what they meant.

This matters beyond voice notes. A route question, a family agenda, or a medicine request can all contain a small mixed-language detail that carries the action. What Happens When Your Twi Question Comes Mixed With English? explores that everyday pattern more closely.

Give people a way to correct the conversation

No voice system should pretend it heard perfectly when it did not. A useful tool shows errors clearly and gives the speaker room to repair the message.

With Nkomo, people can talk or type in natural Twi and Ghanaian English, including the mix they already use. In hands-free voice mode, they can hold to speak, interrupt naturally, and hear a reply. If a phrase needs correcting, the conversation can continue from that point instead of forcing a full restart.

The same principle applies to privacy. A voice note can contain family health details, money matters, or a private instruction. People deserve a clear choice about whether audio reaches the cloud: never, ask each time, or for the current session. They can also control on-device history, turn it off for immediate purging, export their data, or delete their account in one tap.

Akosua did not need a grand lesson about language technology. She needed to send the medicine detail again in a form her cousin could act on. This time, she typed the key phrase and added a photo of the box. Her cousin answered with the right one.

The useful standard is simple: before trusting a transcript, check the words that carry the decision. Names, medicines, places, times, and the Twi phrase someone chose for a reason all deserve a second look.