A Twi caller can stay in Twi when the agent’s persona explicitly sets Twi as the conversation language and the first message opens in Twi. The configuration should also tell the agent how to handle English terms that belong in the task, such as an order number or account reference, without treating them as a reason to change the whole conversation.
At 4:45 p.m. in Osu, an illustrative support lead called Ama is holding a phone beside a paper queue of unresolved delivery questions. She has already heard the same problem twice that afternoon: a caller begins comfortably in Twi, then the automated assistant answers in English after hearing an order reference.
The risk is more than an awkward greeting. The caller may give up before the issue is captured, and Ama’s team will have one more unresolved case to call back manually. With the desk closing soon, there is no room for a voice agent that treats one English term as permission to abandon the caller’s language.
Start with a persona that protects the caller’s language
In Asenda Talk, the persona is where you define the assistant’s role, tone, and language behaviour. For this use case, the instruction needs to be plain enough that the assistant has one job: conduct the support conversation in Twi.
A useful starting point could read:
“You are a calm customer support assistant for a delivery team. Speak Twi throughout the conversation. Keep questions and explanations in Twi. You may repeat order numbers, names, phone numbers, addresses, dates, or product names exactly as the caller gives them. Do not change the conversation language because a caller uses English terms inside a Twi sentence. If you do not understand a detail, ask for it again in Twi.”
That final instruction matters. A caller who says an English order code inside a Twi sentence has not necessarily asked for English support. The agent should preserve the code, then continue with the question in Twi.
Keep the persona narrow. Avoid adding a long list of policies, product facts, escalation rules, and campaign instructions before you have checked the basic interaction. Early access is the right time to test one conversation path at a time: greeting, issue capture, clarification, and a clear handoff when the agent cannot complete the task.
Asenda Talk’s Twi recognition and synthesis are fine tuned in house. The assistant runtime is Vapi orchestrated, while the language behaviour still depends on the instructions you give the persona and the speech cases you evaluate.
Make the first message a real Twi opening
The first message sets the language expectation before the caller has to correct the system. It should sound like the start of the task, not like a generic welcome screen.
For Ama’s delivery support example, she could configure:
“Maakye. Me din de Akosua, na me boa wo ma wo delivery ho nsɛm. Sɛ wopɛ sɛ woka wo order number anaa wo din a, ka kyerɛ me.”
The exact wording should be reviewed by a fluent Twi speaker for the team’s audience and service style. What matters in the configuration is the intent: greet in Twi, identify the assistant’s purpose, and invite the first piece of information in Twi.
Do not make the first message too broad. “How can I help?” leaves the caller to guess what the agent can actually do. “Tell me your order number or name” gives the caller a next step. It also creates a useful first test: can the agent hear a Twi answer that contains an English order reference, then ask its next question in Twi?
Test the point where calls usually drift
Ama’s test case should include a caller saying something like: “Me order number no yɛ AB 47, na mennyaa delivery no.” The agent needs to retain “AB 47,” understand that the delivery was not received, and reply in Twi.
This is where a persona earns its place. If the instruction only says “support customers in Twi and English,” the model has room to choose either language. If the goal is a Twi first interaction, say so directly. English can remain available for exact identifiers or a caller who clearly asks to continue in English.
Test slowly spoken greetings, repeated details, names, and a caller who switches only one phrase into English. What Happens When a Twi Caller Repeats a Greeting More Slowly? covers why that first recovery moment deserves its own evaluation.
Record what actually happened in the call trace: the opening message, the caller’s words, the detected language behaviour, and the outcome. Asenda Talk includes telephony lifecycle webhooks and call truth tracking for this purpose. Consent, opt out, and audit records should be part of the design whenever calls move beyond a controlled test.
Keep the rollout boundary visible
When Ama runs the configuration again, the difference is concrete. The assistant greets in Twi, takes the order reference without turning it into an English conversation, and asks the next delivery question in Twi. Ama can review the trace and decide whether that behaviour is ready for the next test.
That is the useful threshold for this configuration: observed Twi continuity in the cases your callers are likely to bring, rather than a broad claim about every possible conversation.
Asenda Talk is in active early access. Outbound calling remains gated until an explicit telephony provider decision is made live, so this is a configuration and evaluation workflow, not a live outbound calling promise. Before any rollout, test consent language, opt out handling, escalation, and language switching deliberately. Can Your Twi/English Agent Handle a Language Switch Before a Live Calling Rollout? is a useful next check once the Twi first path is working.
Comments
No comments yet.