What Happens When a Caller Switches Between Twi and English?

A credible Ghanaian voice-agent evaluation has to test Twi-English code-switching during a real task, because that is where an English-first demo can lose the caller’s intent. Asenda Talk is being built around in-house Twi speech recognition and synthesis, with configurable agents and a Vapi-orchestrated runtime; it remains in early access, and live outbound calling awaits a telephony-provider decision.

At 4:40 on a Thursday afternoon in Accra, Kwame has his yellow notepad open beside a cooling cup of tea. He is reviewing a voice-agent demo for a customer-support campaign due to be signed off the next morning. The agent handles its opening in English cleanly enough. Then Kwame plays the part of a caller who changes languages halfway through the request, adds a detail in Twi, and asks what happens next.

The first agent keeps moving, but the conversation has already slipped. It catches fragments, answers a different question, then returns to its English script. Kwame tries the same prompt again. The result changes, but the problem does not: the caller’s intended next step is no longer clear.

That creates a real risk for the campaign. A call that sounds polished in the first ten seconds can still turn a payment question, support request, or callback request into a manual follow-up queue. If the team signs off based on the opening alone, they may discover the gap only after callers have already had the experience.

With the decision still sitting on the edge of Friday’s approval, Kwame tests a Twi-first agent configured in Asenda Talk. The conversation stays on the task when the caller moves between Twi and English. The agent can be given a persona, first message, and voice before the test. The evaluation is no longer about whether the greeting sounds acceptable. It becomes about whether the agent preserves meaning through the part of the call where a real person stops following the script.

A familiar accent does not prove language understanding

“Ghanaian English” can describe a voice, a pronunciation style, or a market focus. None of those labels prove that an agent can recognize Twi accurately, speak it naturally, or hold its place when a caller switches languages mid-sentence.

That distinction matters because callers do not announce a language transition. A caller may start in English, use Twi to explain the part that matters, then switch back to English to ask for action. If the system misses the Twi detail, the response may remain fluent while becoming wrong.

A useful demo therefore includes a task with consequences. Ask the agent to capture a name, clarify a request, confirm a next step, and respond when the caller changes language. Listen for more than pronunciation. Check whether the agent understood the full request and whether its answer still matches it.

The demo should also include difficult but ordinary moments: a correction after the agent misunderstands, a short interruption, or a caller who changes their mind. Those are the moments that reveal whether the conversation is being carried forward or merely transcribed one phrase at a time.

For a deeper testing framework, see this Twi voice-agent demo checklist.

Build the evaluation around the caller’s intended next step

Kwame’s second test is short on purpose. He gives the agent one job: understand what the caller wants to happen after the conversation. That could be a callback, a support handoff, a payment follow-up, or a request to stop being contacted.

This is a stronger test than asking, “Does it speak Twi?” A voice agent earns trust when the caller’s next step remains intact after a language switch, clarification, or interruption.

Write five to ten prompts that reflect the calls your team expects. Include your actual terms, names of services, and the Twi-English phrasing your staff hear. Then score each test against a plain standard:

• Did the agent identify the request?
• Did it retain the detail after the caller code-switched?
• Did its response make the next action clear?
• Could an operator later verify what happened on the call?

The last question matters when a campaign is tied to real work. A transcript or call status that only says “started” does not tell an operations lead whether the caller got a useful outcome. Asenda Talk’s telephony lifecycle webhook pipeline is designed for call-truth tracking, so teams can connect call events to what actually occurred rather than treating a started call as a completed conversation.

Treat safety and spend as part of the product test

Language performance is only one part of a voice-agent evaluation. A business also needs to know what happens when a caller says they do not want further contact.

In Kwame’s evaluation, he adds that request near the end of the call. He wants to see that consent, opt-out handling, and the audit trail are considered before anyone plans a campaign. An agent that recognizes Twi but cannot support responsible contact practices creates a different operational problem.

Asenda Talk includes consent, opt-out, and audit-trail capabilities for every call, along with metered per-minute billing and an operator-controlled real-money gate. Those controls are practical for teams testing an early-access platform: they can inspect configuration and cost exposure before turning a test into a financial commitment.

The platform also keeps administrative secrets write-only, masked, and environment-aware. That detail rarely makes a demo reel, but it matters when a technical team moves from an evaluation into a controlled deployment.

Make Friday’s decision smaller and more honest

By the end of the afternoon, Kwame does not need to claim that one demo settles every question. He needs enough evidence to make the next decision responsibly.

He records which prompts held their meaning, where the agent needed improvement, and which parts of the intended workflow are available today. He also separates the current platform from the roadmap. Asenda Talk can support building and configuring voice agents, native Twi speech work, runtime orchestration, call-truth infrastructure, billing controls, and call-safety records today. It does not yet offer live outbound calling until its telephony-provider path is decided and made available.

That clarity gives the team a better next step than a broad promise. They can test the exact conversations their callers will have, keep a record of what worked, and avoid approving a campaign on the strength of an English greeting alone.