A Ghanaian voice agent must carry the customer’s intent, prior answers, and call state across an English-to-Twi switch. Translating each sentence alone can lose the meaning of the conversation, especially when consent, a deadline, or an opt-out depends on what was said before the switch.
At 4:47 p.m., an illustrative shop owner named Efua stood behind her counter in Kumasi, folding a small receipt while two customers waited near the door. An agent had called in English about a delivery follow-up. Efua confirmed the order number, then switched to Twi as she explained that the recipient would be unavailable and asked for the date to be changed.
The risk was already on the table. If the agent treated the Twi sentence as a fresh, disconnected request, it could confirm the wrong change, fail to record the customer’s preference, or continue a conversation after she had asked it to stop. Efua had no reason to repeat every detail in English just to make the system understand.
The useful turn comes when the agent retains the thread: the order already discussed, the contact’s availability, the reason for the call, and the customer’s instruction. The language changes. The task does not.
A language switch carries more than vocabulary
English and Twi often share the same call because that is how people actually speak. A customer may begin with the language used in the first message, move into Twi when the details become personal or urgent, then return to English for an account number or product name.
That switch is full of context. “Change it” only makes sense if the agent still knows what “it” refers to. A Twi response that confirms a time, rejects an offer, or asks for a callback needs to connect to the earlier English turns. The agent also needs to preserve who is speaking, why the call began, and what action is allowed next.
This is particularly important in calls where the customer’s words change the call’s status. A request to stop calling, for example, should not become a vague note in a transcript. It needs to trigger the right opt-out handling and appear in the audit trail. The same applies to consent, appointment changes, verification responses, and billing-related questions.
A translated transcript can be useful for review. It cannot be the whole system of record.
Context is the difference between a conversation and a sequence of clips
Efua’s call has a simple job: update a delivery arrangement. But a real agent must track several things at once.
It needs to know the call’s purpose and the details already collected. It needs to recognize when the customer changes language. It needs to keep the customer’s instruction attached to the right record. And it needs to preserve the lifecycle events that show what happened during and after the call.
That is why voice design starts before speech recognition. The agent’s persona, first message, prompts, call states, escalation paths, and exit conditions should all account for code-switching. A prompt that assumes the customer will stay in one language creates friction at the exact point a customer is trying to be clear.
Asenda Talk is being built for this problem with native Twi speech recognition and synthesis fine-tuned in-house. The platform also lets operators configure an agent’s persona, first message, and voice, while its telephony lifecycle webhook pipeline tracks call events. The assistant runtime is Vapi-orchestrated.
Those components still need evaluation in the calls that matter to a business. Early-access teams should test English-to-Twi switches against their own scripts, terms, names, place references, and escalation rules. A language model can recognize words correctly and still mishandle the underlying task.
Test the moments where the customer changes the call
A useful test plan does not begin with a polished demo. It begins with the messy middle of a real conversation.
Create test calls where a customer confirms an identity detail in English, switches to Twi to correct it, then asks in English what happens next. Test a callback request after a language switch. Test a refusal. Test an opt-out. Test the case where the customer speaks partly in Twi and partly in English within one answer.
For each test, review more than the transcript. Check whether the agent’s next response matches the full conversation. Check that the webhook events show the right call state. Check that consent and opt-out records are present and usable in an audit. If billed minutes are involved, reconcile the call events with the meter before allowing real charges. The operational side matters as much as the voice quality, as this review of disputed calls explains.
Efua’s scenario also needs a safe fallback. If the agent cannot confidently preserve the request across the switch, it should state that clearly, collect the minimum needed detail, and route or pause the task according to the operator’s process. Guessing is a poor substitute for comprehension.
Build trust before live outbound calls
Natural Twi and English conversation is a product goal. It does not remove the need for control over when, where, and how calls happen.
Asenda Talk includes consent, opt-out, and audit-trail handling for calls, plus per-minute metering with an operator-controlled real-money gate. Outbound calling remains gated behind an explicit telephony-provider decision that has not been made live. That constraint is deliberate: a convincing voice should never be mistaken for permission to launch a campaign.
When Efua ends the call, the useful outcome is modest and concrete. Her delivery change is attached to the right conversation, her language choice did not make her repeat herself, and the operator has a record to review if something later needs checking.
That is the standard worth testing: after the customer switches languages, can the agent still act on what the customer meant?
Comments
No comments yet.