A claim that an AI “supports Twi” only tells you that Twi may appear somewhere in its input or output. For a mixed sentence, the useful test is whether it keeps the meaning, tone, and turn-taking intact when English and Twi arrive together.
Use a sentence people would actually say, such as: “Mepa wo kyɛw, tell Ama sɛ meeting no afi 3, but don’t make it sound harsh.” It asks the AI to understand Twi politeness, English instructions, a named person, a time, and the tone of a message.
Start with the meaning, not the language label
A system can handle isolated Twi words and still struggle with the sentence as a whole. “Mepa wo kyɛw” carries a polite request. “Meeting no afi 3” means the meeting starts at 3. The English ending changes the task again: write the message gently.
If an AI treats each language as a separate translation problem, it may translate phrases one by one, miss the relationship between them, or reply in a language mix that feels unnatural. It may also flatten the request into a literal translation and lose the social instruction behind “don’t make it sound harsh.”
Code-switching support means the model follows the full thought across both languages. The language boundary should not become a meaning boundary.
Compare the same prompt in four tools
Run the exact sentence in text first, then in voice if the product offers it. Ask each tool to produce the final message for Ama, then check four things:
- Did it understand that the meeting begins at 3, rather than ending then?
- Did it keep “Mepa wo kyɛw” as a polite request rather than a word-for-word fragment?
- Did it follow the instruction about sounding gentle?
- Did it preserve the mix of Twi and Ghanaian English in a way you would actually send?
Nkomo is built for natural Twi and Ghanaian English conversation by text or voice, including ordinary code-switching. For this prompt, the practical expectation is that you can speak or type the whole mixed sentence as one turn, receive a concise reply, and correct or interrupt it naturally in hands-free voice mode. Its privacy choices are also explicit: you choose whether cloud use is never allowed, asked each time, or allowed for the session. If you turn off on-device history, it is purged immediately.
Khaya is associated with GhanaNLP and local-language AI, but a product name alone does not establish how its current interface handles this particular mixed instruction. Test it directly. Check whether it accepts the entire prompt without requiring you to separate Twi from English, and whether it preserves the intended tone in its answer.
Gemini and ChatGPT Voice can be useful comparison points because they are general-purpose AI assistants with broad language capabilities. Their Twi and voice behaviour can change with product updates, device settings, region, model selection, and the wording of the prompt. Treat a good single result as a result to repeat, rather than proof of consistent code-switching support.
That caution matters. GSMA’s emphasis on local-language AI and local data ownership points to two separate questions: can the system understand how people speak, and can people see what happens to their data? A fluent answer does not answer the second question.
Test speech separately from text
A typed prompt gives the system clean spelling and punctuation. Voice adds accents, pace, background noise, interruptions, and the everyday rhythm of switching languages mid-sentence.
For voice testing, say the sentence naturally. Do not slow down into a classroom pronunciation exercise. Try a second version with a small interruption: “Mepa wo kyɛw, tell Ama sɛ meeting no afi 3... wait, make it 3:30.” A useful voice assistant should follow the correction and make the current time clear.
Listen for more than pronunciation. Did it hear “afi” correctly? Did it confuse 3 with 3:30 after your correction? Can you interrupt its response before it continues with the wrong answer?
Nkomo’s hold-to-speak mode is designed for this back-and-forth. It also shows errors rather than swallowing them, which helps you distinguish a failed request from an answer you can rely on. For any other tool, note whether an error message, transcript, or retry path tells you what went wrong.
Keep a small scorecard
Use three prompts rather than one. Add a family message, a practical request, and a sensitive instruction. For example:
- “Bra ha kakra, but text me first because I’m in a meeting.”
- “Please explain this bill in simple Twi, then give me the total in English.”
- “Mma no nka sɛ mekae, just say I may be late.”
Score each answer for meaning, tone, natural language mix, correction handling, and privacy clarity. A tool that translates a sentence correctly but makes the final message rude has failed the task. A tool that speaks well but gives unclear data controls leaves another important decision unresolved.
Choose the tool based on the whole conversation
If your priority is natural Twi and Ghanaian English conversation with clear controls over cloud consent and local history, test Nkomo with your own mixed prompts. If you are comparing it with Khaya, Gemini, or ChatGPT Voice, use the same wording, the same voice conditions, and the same scoring sheet.
Then send one message you genuinely need to send. Read it aloud before you use it. If it sounds like something you would never say to Ama, ask again, more directly, or choose a tool that handles the mix with more care. For a deeper look at what can be lost when a system flattens tone, read What Happens When a Transcript Flattens a Twi Rebuke?.
Comments
No comments yet.