Conversational AI should follow a speaker’s meaning across street noise, rushed phrasing and an English-to-Twi turn. Requiring a clean language choice before every sentence breaks the conversation people are already having.
When the message does not arrive neatly
In April 1970, Apollo 13’s crew faced rising carbon dioxide while the spacecraft was far from Earth. The command module had square lithium hydroxide canisters. The lunar module needed round ones.
At NASA’s Mission Control in Houston, flight controller Ed Smylie and his team had to make those incompatible parts work using materials already aboard the spacecraft. NASA’s account describes the improvised adapter they developed and relayed to the crew, who built it under pressure.
The problem was larger than the shape of a canister. Two systems involved in the same mission could not understand each other at the moment understanding mattered most.
A voice note recorded outside a provision shop carries a quieter version of that problem. A customer may begin in English, rush through the request over passing traffic, then switch to Twi for the detail that matters: “Please tell Kofi I’ll come later, na ɔntwɛn me.”
The meaning belongs to one person and one intention. A language selector can split it into separate inputs and lose the relationship between them.
Track the speaker through the switch
Natural conversation does not pause for configuration. People repeat themselves, interrupt, change language for emphasis and use whichever expression reaches the point fastest.
That is common in Ghanaian speech, and it appears in research data too. KasaSpeech contains 54,855 manually transcribed recordings, more than 95 hours, from speakers across Ghana. The collection covers natural English and Twi switching in everyday topics and communication scenarios.
The useful design question is therefore not, “Which language did the user select?” It is, “What is this person trying to say across the whole turn?”
That means retaining enough conversational context to connect the English opening with the Twi correction, even when the audio includes background noise or the speaker is hurrying. It also means showing an error when the system cannot follow, instead of quietly returning a polished but wrong answer.
This is why Twi and English must work in the same sentence, and why a small word such as “nanso” can change the person a message refers to.
Let speech behave like speech
Nkomo lets people talk or type in natural Twi, Ghanaian English or a mix of both. In voice mode, you hold to speak, hear the reply and interrupt naturally when you need to correct or add something.
The goal is simple: keep pace with the person instead of asking the person to reorganize their speech for the software.
Privacy controls remain explicit while the conversation moves. You choose whether cloud use is never allowed, requested each time or allowed for the current session. History stays on the device under your control, and turning it off purges it immediately. You can also export your data or delete your account with one tap.
Those controls matter because understanding a mixed-language turn requires context, and people deserve a clear say in where that context goes and how long it stays.
Test the difficult turn, not the clean demo
A useful voice test should include the conditions that expose weak assumptions: background sound, a hurried opening, one interruption and a language switch carrying an important detail.
Then check whether the response preserves the speaker’s intention from beginning to end. If the system misses the turn, it should say so clearly. A visible error gives the person a chance to try again. A silent failure can send the wrong message with confidence.
Smylie’s team did not ask Apollo 13’s crew to make the square canister round. They built a connection between the systems already in use. Conversational AI for Ghana needs the same discipline: meet the speaker where the meaning is, including the moment English becomes Twi outside the provision shop.
Comments
No comments yet.