The Creased Receipt Kojo Held, and the Second Charge It Could Not Rule Out

A bilingual voice agent should pass one simple test: give it a single request that begins in Twi and ends in English, then check whether its response preserves the meaning of the whole utterance. If it handles each language fragment separately, a clean transcript can still produce the wrong action.

At 4:47 p.m. in Accra, Kojo is holding a creased payment receipt while his phone battery sits in the red. He tells a support agent, “Mepɛ sɛ metua ka no, but don’t charge me twice.”

Kojo is an invented composite, but the failure in front of him is plausible. The Twi half says he wants to pay the bill. The English half adds the condition that matters: avoid a second charge.

If the agent hears only “I want to pay,” it may send him back into the payment flow. If it hears only “don’t charge me twice,” it may classify the call as a duplicate-payment dispute without understanding that he has not completed the new payment. Either route could leave the account unresolved and put another charge at risk.

His battery drops again. There may be no second call.

One sentence should produce one meaning

Code-switching is ordinary conversation for many bilingual speakers. A person can establish an action in Twi, qualify it in English, then expect the listener to treat both parts as one request.

That expectation gives testers a useful diagnostic. Keep the utterance short enough to inspect, but make the second half change the meaning of the first:

“Mepɛ sɛ metua ka no, but don’t charge me twice.”

The test is stronger than asking the agent one question in Twi and another in English. Separate questions can be routed through separate language paths and still look successful. A mixed sentence forces the speech system and assistant runtime to preserve intent across the switch.

Listen for the agent’s next move. Does it confirm both the intended payment and the duplicate-charge concern? Does it ask a relevant question before taking action? Or does it repeat one fragment and lose the relationship between them?

The response should demonstrate understanding in plain language. A polished voice cannot compensate for a broken intent.

Inspect the boundary, not only the transcript

A transcript can look tidy while hiding the failure that matters. The words may all appear on screen, yet punctuation, timing, or language segmentation can cause the assistant to treat them as two unrelated inputs.

Test the full path:

• Record the mixed utterance exactly as spoken.
• Compare the audio with the recognized text.
• Inspect where the system marks the language change.
• Check the intent passed into the assistant runtime.
• Review the response and any action the agent attempted.
• Confirm that consent, opt-out events, and call outcomes remain attached to the same call record.

This matters because a bilingual failure can surface after speech recognition. The recognizer may preserve every word while the runtime assigns the wrong intent. It can also happen earlier, when the final English condition disappears or gets attached to the next turn.

The same principle appears in What Happens When a Twi Request Disappears From an English Transcript?: readable output does not prove that the system preserved the caller’s request.

Change the second half and watch the action change

Run the sentence again with one controlled variation. Keep the Twi opening stable and replace only the English ending:

“Mepɛ sɛ metua ka no, but I already paid this morning.”

Now the safe next step should change. The agent should investigate the earlier payment before inviting another one.

Try another ending:

“Mepɛ sɛ metua ka no, but call me next week.”

That version adds a timing instruction. If the agent immediately continues the payment process, it has recognized words without following the caller’s constraint.

These pairs reveal more than a general bilingual demo. They show whether the system carries meaning across the boundary and whether a small language change produces the correct operational change. Keep the audio, transcript, intent, response, and call-truth events together so reviewers can trace the failure without guessing.

Asenda Talk supports native Twi speech recognition and synthesis fine-tuned in-house, with Vapi orchestrating the assistant runtime. The platform is in active early access. Voice-agent configuration, call lifecycle tracking, consent, opt-out records, audit trails, and operator-gated per-minute billing are built today. More African languages remain in progress, and live outbound calling still depends on an explicit telephony-provider decision.

That distinction should shape testing. Evaluate the language path that exists now. Do not treat a configured agent as proof that live outbound delivery is ready.

Give Kojo a response he can safely act on

Back at 4:47 p.m., the useful turn arrives when the agent reflects the complete request: Kojo wants to pay, and he needs confirmation that the payment will not be duplicated. It asks for the information required to check before moving him forward.

The payment outcome is still unresolved, as it should be. The agent has not invented certainty or pushed him into another transaction. It has preserved the condition that protects him.

Before a bilingual agent reaches real calls, add one mixed-language sentence to every critical intent test: payment, cancellation, consent, opt-out, delivery, identity checks, and escalation. Begin in Twi. End in English. Then verify that the action reflects the whole sentence.

Kojo can put the receipt back in his pocket. His battery may still be low, but the next step is finally the right one.