A capable Twi voice agent should preserve intentional English account terms when the caller uses them. Translating every word can change the meaning, especially when a familiar banking label such as “savings account” refers to a specific product in the caller’s mind.
At 4:40 on a humid afternoon in Kumasi, Abena stood behind the counter of her small fabric shop, folding a length of blue print while answering a support call. This is an illustrative composite, designed to show a common bilingual interaction rather than describe a real customer.
She spoke mostly in Twi. Then she reached the detail that mattered: the transfer had come from her “savings account.”
The agent translated the term instead of preserving it. Its next question treated her words as a general description of money she had saved, rather than the name she used for the account. Abena corrected it once. The agent translated again.
Now the call record risked attaching the disputed transfer to the wrong account. If that distinction disappeared before review, her report could be delayed or classified incorrectly. She put down the fabric and repeated two words in English: “savings account.”
This time, the system kept them.
Borrowed words can carry the precise meaning
A Twi conversation does not become less natural because it includes an English banking term. Ghanaian speakers may move between languages within one sentence, especially when discussing account types, mobile services, workplace processes, or product names.
That switch can be deliberate. The caller may know a service by its English label because that is how it appears in an app, on a statement, or in previous conversations with staff. Replacing the label with a literal translation can make the transcript sound tidier while making the underlying record less accurate.
The right question is not, “Can this word be translated?” It is, “What did the caller intend this phrase to identify?”
For Abena, “savings account” was an identifier. It separated one account from another. Preserving the phrase protected that distinction in the transcript and gave the next step in the call a reliable reference.
The same issue appears when a caller gives one English answer inside a Twi exchange. A single borrowed word can carry more operational meaning than the surrounding sentence.
Recognition must follow the conversation, not a language toggle
A bilingual call rarely waits for a clean boundary between Twi and English. The caller may begin in English, explain the difficult part in Twi, and insert an English product term halfway through that explanation.
A system built around a rigid language setting can struggle here. If it assumes that every sound inside a Twi turn must become Twi text, intentional borrowing may be translated, distorted, or dropped. If it treats the turn as English, the surrounding Twi may suffer instead.
Asenda Talk approaches this problem through native Twi speech recognition and synthesis fine-tuned in-house, with Vapi orchestrating the assistant runtime. The practical goal is to evaluate the words the caller actually chose, including deliberate switches, rather than force the whole turn into one language.
That work remains part of active early access. A useful evaluation should therefore examine more than whether the transcript looks fluent. It should check whether account labels, names, amounts, consent language, and corrections survive the switch unchanged.
This matters most at the hardest point in the call. As discussed in what happens when a support call switches to Twi, language choice often changes when the caller needs precision, reassurance, or room to explain.
Corrections need to survive beyond the audio
Once Abena repeated “savings account,” the correction needed to affect the record. A voice agent that acknowledges her verbally but leaves the earlier interpretation untouched has only performed politeness.
A dependable call flow should retain what was heard, what the system inferred, and what the caller corrected. That evidence matters when a support worker reviews the call or when an operator investigates why the agent chose its next question.
Asenda Talk includes a telephony lifecycle webhook pipeline with call-truth tracking, alongside consent, opt-out, and audit records for each call. Those controls create places to verify what happened during an interaction. They do not guarantee that every bilingual term will be recognized correctly, particularly during early access. They make errors and corrections easier to inspect instead of hiding them behind a polished final transcript.
For teams evaluating a Twi voice agent, the test should include realistic borrowed terms spoken inside full Twi sentences. Use the account labels, service names, and internal phrases callers already say. Then inspect the transcript, the agent’s interpretation, its follow-up question, and the audit trail.
Test the words that must remain untouched
Abena’s call ends differently once “savings account” stays in English. The next question refers to the correct account, and the record keeps the phrase she chose. She picks up the blue fabric again, knowing the report at least identifies the right place to investigate.
That is the standard worth testing: the system preserves meaning when the speaker changes language for a reason.
Teams can start with a short evaluation set. Put one intentional English term inside each otherwise Twi exchange. Include a correction after an initial misunderstanding. Check whether the agent updates its interpretation and whether a reviewer can trace that change later.
Asenda Talk can be configured with an agent persona, first message, and voice for this evaluation. Live paid outbound calling remains gated while the telephony-provider decision is unresolved, so testing should not be presented as an approved outbound campaign. The immediate task is narrower and concrete: find the terms your callers mean to keep, then verify that the agent leaves them intact.
Comments
No comments yet.