Ama’s voice note was interrupted. The envelope had to leave that evening.

People speaking Twi and Ghanaian English should be able to continue one thought without choosing a language before they speak. Code-switching is common in contemporary spoken Twi, and an assistant that interrupts to demand a language setting creates work the speaker never asked for.

At 6:42 p.m., Ama was standing outside her auntie’s house in Kumasi with a phone in one hand and a folded blue envelope in the other. Her cousin was leaving soon, carrying documents that needed to reach the family before morning. Ama pressed record and began a voice note: “Auntie, the envelope no, fa ma Kwaku. The photocopy bi wɔ mu, but please tell him sɛ ɔnhwɛ page last no.”

Halfway through, the AI paused her.

“Which language would you like to use?”

Ama stared at the screen. She had already chosen. She was using the language she uses with her family.

The question sounds small until the message has a consequence. If Kwaku received an incomplete instruction, the envelope could leave without the page the family needed. Ama could type everything again in formal English. She could record separate messages, one in Twi and one in English. She could call, hope someone answered, and repeat herself over traffic noise.

For a moment, none of those options solved the problem in front of her.

A language setting can break a real conversation

People do not always divide a thought into neat language turns. A sentence can start in English, switch into Twi for emphasis or intimacy, then return to English for the detail that must be clear. That mix carries meaning. It can show who the message is for, how urgent it feels, or which phrase sounds right in the family.

A forced language choice treats that movement as an error to correct. The speaker has to translate their own life into the structure an assistant expects.

That is a poor trade when the speaker is already trying to think quickly.

The cost may be small in a casual chat. It grows when someone is explaining a recipe while their hands are wet, checking a message for an older relative, or recording instructions before a trip. The interruption adds a decision at the exact point the person came for help.

The problem can get worse when the assistant goes quiet after the switch. A spoken interaction has momentum. Once that momentum breaks, people do not know if their words were heard, misunderstood, or lost. What happens when a voice assistant falls silent at an English-Twi switch? explores what that uncertainty costs when a conversation needs to keep moving.

The speaker already made the choice

Ama did not need an assistant to tell her whether her message was Twi or English. She needed help with the message as she spoke it.

That is the design idea behind Nkomo. You can talk or type in natural Twi and Ghanaian English, including the ordinary switches people make mid-thought. In voice mode, hold to speak, interrupt naturally, and hear the reply without turning the conversation into a settings exercise.

For Ama, the turn comes when she records the message in the words she had already chosen. She can ask for the instruction to be made clearer without flattening it into a language that changes its tone. The envelope still leaves that evening, and Kwaku receives a message that tells him which page to check.

The morning after is less dramatic. Ama is back at work, her phone is quiet, and she has not spent the night wondering whether a menu prompt caused the family to miss an important document.

That is the point. Good language support often looks like fewer interruptions.

Privacy needs the same plain-spoken clarity

Language choice is only one place where an assistant can take control away from the person speaking. Privacy can create the same feeling when a voice message leaves the device without a clear, timely choice.

A family voice note may include names, money worries, health concerns, or an argument that should stay within the family. People deserve to know when information reaches the cloud and what remains on their device.

Nkomo makes that choice visible. You can set cloud consent to never, ask each time, or this session. You can turn off on-device history, which purges it immediately. You can also export your data or delete your account in one tap.

Clear controls do not require a person to become a privacy expert. They give someone like Ama a direct answer before she speaks: what will be sent, what will stay local, and what can be removed.

Build for the sentence people actually say

A conversational assistant should meet people in the language pattern they bring to the moment. That means recognising a switch as part of the message, not a failure of the speaker to stay in one lane.

It also means being honest when something goes wrong. Nkomo shows errors rather than swallowing them silently, so a person can decide what to do next instead of guessing whether a message disappeared.

Ama’s blue envelope was never really about a language menu. It was about being understood while she was trying to help her family, in the voice that came naturally.