AI for Ghanaians should handle English and Twi as they appear in one thought, one voice note, and one turn. The 54,855 transcribed recordings in KasaSpeech’s 95-hour English-Twi code-switching dataset point to a practical design requirement: treat switching as normal speech, then make it clear when that speech reaches the cloud.
A voice note does not announce its language changes
A family voice note may begin with, “Please tell Auntie I’ll call her after work,” move into Twi for the part that needs warmth or emphasis, then return to English for a time, place, or plan. Nobody pauses to label the change. Nobody says, “I am now switching languages.”
That is the interaction an assistant has to meet.
A person recording the message is usually trying to get something done. They may be coordinating a pickup in Kumasi, checking on someone in Accra, or replying from London after a long day. The language mix carries meaning as well as information. Twi can signal closeness, urgency, correction, humour, or the exact phrasing a relative will recognise. English may carry the logistics. Separating the two into neat language turns can make the conversation feel unlike the way the speaker actually speaks.
For AI, this means the useful unit is often the whole utterance, not a language label attached to each sentence. A system needs to follow the thread when “I’ll send it later” becomes “na me frɛ wo,” and when the important instruction arrives halfway through a mixed reply.
That matters most when the message affects money, care, travel, or family expectations. A missed word is frustrating. A misunderstood instruction can be costly. What happens when the important instruction arrives in Twi? explores why the language carrying the key detail cannot be treated as an optional extra.
The Apollo 13 fix had to fit what was already there
In 1970, Apollo 13 faced a carbon dioxide problem after an explosion damaged the spacecraft on its way to the Moon. The lunar module had a limited supply of lithium hydroxide canisters, while usable canisters in the command module had a different shape. The problem in Houston was immediate: the crew needed a way to use square canisters with a round opening.
Ed Smylie led the engineering effort that devised an adapter from materials available aboard the spacecraft. The crew assembled it, and the solution worked. NASA’s Apollo 13 mission archive documents the crisis and the improvised carbon dioxide scrubber fix.
The point was not to make every component identical. The point was to preserve the connection between systems that were already different, under pressure, with no room for a vague handoff.
English and Twi in a Ghanaian voice note are not faulty inputs that need tidying before an assistant can respond. They are the shapes the conversation already has. A useful assistant has to work with those shapes: hear the switch, retain the thread, and respond in language that fits the speaker’s own mix.
Apollo 13 also offers a quieter lesson. The team could see the problem. They did not receive a reassuring blank screen while carbon dioxide continued to rise. For voice AI, visible errors matter for the same reason. If something fails, the person speaking needs to know, especially when they are relying on the answer to act.
Data should reflect speech, then product choices should respect people
A dataset of 54,855 transcribed recordings gives builders evidence that mixed English-Twi speech is substantial enough to design for directly. It can help test whether a system loses meaning at a switch, mistakes a Twi phrase for a name, or replies as though the speaker had chosen one language and stayed there.
But language coverage and privacy are separate promises. A product can recognise that a conversation moves naturally between English and Twi while still being vague about where the recording goes, how long it stays, or whether the user can remove it. That vagueness asks people to trust a process they cannot see.
Nkomo makes those choices explicit. You can speak or type in natural Twi and Ghanaian English, including the mix that comes naturally in conversation. In voice mode, you hold to speak, can interrupt naturally, and hear the reply. Before cloud use, you choose whether to never allow it, be asked each time, or allow it for the session. History stays on the device unless you choose otherwise, and turning it off purges it immediately.
Those controls matter because a voice note can contain more than a question. It can contain a family matter, an address, a payment instruction, or somebody’s health update. The right question is not only, “Can the assistant understand this?” It is also, “Do I understand what happens to it?”
Build for the whole conversation
The most useful test for a Twi-English assistant is simple: give it a voice note someone would actually send, with the ordinary switches left in. Check whether it follows the meaning, keeps the important instruction intact, and responds without forcing the speaker into a language menu.
Then check the consent path before the first sentence. A person should be able to see the cloud choice, control local history, export their data, and delete their account without hunting through settings. Why the cloud choice must come before the first sentence is a useful standard for that part of the experience.
Apollo 13’s engineers did not solve their problem by asking the crew to reshape every square canister into a round one. They built the connection around the real parts available. AI designed for Ghanaian speech should begin there too: with the voice people already use, the language switches they do not notice, and controls clear enough that nobody has to guess.
Comments
No comments yet.