Twi voice assistants need to understand what a person means in the moment, including mixed Twi and Ghanaian English, relationships, corrections and the reason behind a request. Accurate transcription matters, but a useful reply depends on intent and context after the words have been heard.
At 7:40 on a Saturday morning in Kumasi, Efua is holding a phone against her ear while stirring koko with the other hand. Her uncle is expected later, and she wants help putting a short message into words: ask Auntie Mansa whether he is still coming, then remind her to bring the blue folder.
She starts in Twi, switches to English for “blue folder,” then corrects herself. The folder is not for her uncle. It is for her younger brother’s school matter. If the message goes to the wrong person, the folder may not arrive before the family meeting, and Efua will be the one explaining why.
A voice assistant can transcribe every word of that request and still fail the job. It may treat the correction as a new request. It may miss who “she” refers to. It may produce a formal sentence that sounds odd coming from Efua. The difficult work begins after the speech becomes text.
A transcript does not reveal the task
People rarely speak as though they are filling in a form. They begin a thought, revise it halfway through, refer to someone already known in the conversation, and use tone to signal whether they are asking, confirming, joking or warning.
In ordinary Ghanaian speech, Twi and English often sit in the same sentence because each word does a particular job. A person might use Twi for the relationship and English for a school, workplace or phone term. That is not a broken sentence waiting to be corrected. It is the sentence they meant to say.
The assistant has to follow the purpose underneath the phrasing. Is someone asking for a translation, composing a message, checking whether a payment recipient is correct, or simply continuing a conversation? Those are different jobs. A reply that gives a dictionary definition when the speaker needed help wording a respectful message creates more work than it removes.
That is why [forcing people to choose one language]( /blog/what-happens-when-a-voice-assistant-forces-you-to-choose-one-language-e7feb395/) can make a voice interaction feel unnatural before the assistant has even answered. The speaker must first reshape their own language to fit the tool.
Code-switching carries context
A mixed-language request can contain small clues that change the right response. A family title may tell you the tone should be respectful. A correction after a name may mean the speaker is worried about sending something to the wrong person. A pause before “actually” may signal that the first interpretation is no longer safe to use.
Names are a clear example. “Send it to Ama” is simple only when there is one Ama in view. In a real conversation, the speaker may mean Ama from church, Ama at work, or the Ama whose number was mentioned two sentences earlier. A responsible assistant should avoid acting as if an uncertain name is certain.
This is especially important where the outcome has consequences. A wrong reply about a payment name, an appointment, or a family message can turn a small language gap into a costly mistake. The useful behaviour is often modest: ask a short follow-up, repeat the key detail back, or say clearly that the request was not understood.
That last part matters. An assistant that silently guesses can sound confident while moving the conversation in the wrong direction. Nkomo is designed to show errors rather than swallow them, so a person can decide what to say next instead of assuming the system understood.
Voice conversations need room for real speech
Efua does not speak in neat turns while cooking. A pot lid rattles. Her cousin asks where the cups are. She interrupts herself because she remembers the folder is for her brother, then returns to the message.
That is closer to how voice assistance is used at home, in a car, on a walk, or between tasks. The interaction needs to tolerate interruption and allow a person to keep the thread of the conversation. Nkomo’s hands-free voice mode lets people hold to speak, interrupt naturally and hear a reply, rather than forcing a long, polished prompt.
Concise replies help too. When someone is speaking aloud, they should not need to listen through a lecture to find the one sentence they need. The best answer may be a clear draft message, a brief confirmation, or one question that prevents a wrong assumption.
For people practising family conversations, the aim can be different again. They may want a natural exchange that keeps moving when Twi and English appear together, rather than a classroom exercise that stops after every phrase. [This distinction between a learned phrase and a sustained mixed-language conversation]( /blog/practical-twi-practice-versus-natural-code-switching-why-learning-a-family-phrase-and-sustaining-a-mixed-language-conversation-are-different-jobs-cd0b7bd3/) changes what “good” sounds like.
Privacy is part of the conversation
Speech can contain names, family concerns and half-finished thoughts. When a person uses a voice assistant in Twi or Ghanaian English, they deserve to know what leaves the device and what remains on it.
Nkomo makes that choice explicit with cloud-consent controls: never, ask each time, or this session. Its history stays on the device under the person’s control, and turning history off purges it immediately. Data export and account deletion are available in one tap.
Back in Efua’s kitchen, the message is now short enough to send: a respectful check-in for Auntie Mansa, with the folder attached to the right person. The koko has stopped bubbling over. Her next concern is finding the blue folder before her cousin moves it again.
That small outcome is the standard worth building toward. A Twi voice assistant should hear the words, follow the meaning, and leave the speaker in control of the conversation and the record it creates.
Comments
No comments yet.