A voice agent can appear to handle Twi while quietly teaching the caller to switch to English. The failure begins when it misses one familiar phrase, responds with the wrong assumption, and makes English the quickest path forward.
In a composite test call, assembled from the kinds of turns a Ghanaian support desk should test, a caller begins in English and moves naturally between English and Twi. The agent follows until she uses a familiar Twi phrase to signal that she wants the previous point explained again.
The agent does not crash. It does something more misleading: it answers a different question.
There is a pause. The caller repeats herself in English. The agent responds correctly, and the call continues.
On a transcript, that can look like recovery. In practice, the caller has started accommodating the system.
The dangerous call can sound normal
A visible failure is easy to catch. The agent stops speaking, returns an error, or produces an answer that makes no sense.
A missed phrase can hide inside an otherwise orderly conversation. The caller repairs the exchange by simplifying a sentence, changing pronunciation, or switching to English. Once the agent understands, the rest of the call may score well.
That score misses the important event. The human changed languages because the agent could not carry its share of the conversation.
The pattern matters for Ghanaian businesses serving customers who use Twi and English together. Callers may begin with a greeting in English, explain a sensitive detail in Twi, then return to English for a number or name. The hardest part of the call may arrive in the language that feels most natural under pressure, as explored in what happens when the hardest part of a support call switches to Twi.
A test that records only task completion will miss this. The appointment was booked. The account category was selected. The caller reached the final prompt. Yet the agent succeeded only after the caller reduced the linguistic burden placed on it.
One phrase can change what the system believes
On January 25, 1990, Avianca Flight 052 was approaching New York after a long journey from Bogotá. The Boeing 707 had been placed in holding patterns and was running critically low on fuel.
The crew told air traffic control that they were running out of fuel and could not accept further delay. They did not use the explicit emergency declaration controllers expected. The seriousness of the situation was not fully understood in time. The aircraft crashed at Cove Neck, New York.
The US National Transportation Safety Board documented the sequence in Aircraft Accident Report AAR-91/04. Its findings examined fuel management, communication between the flight crew and controllers, and the failure to communicate the emergency with sufficient clarity.
The scale and stakes are entirely different, but the mechanism is relevant. A person can believe they have communicated something important while the receiving system assigns a weaker meaning. The exchange continues, so the misunderstanding stays hidden.
That is the risk in a bilingual voice test. The key question is not whether the system produced words after hearing Twi. The test must determine whether it recognized the caller’s intended action, preserved the relevant detail, and responded without requiring an English repair.
Test the repair, not only the response
Start by marking every point where the caller changes language after an agent response. Treat that switch as evidence to inspect, rather than proof that bilingual conversation worked.
Replay the preceding turn and ask three questions:
- Did the speech recognizer capture the phrase accurately?
- Did the agent map the phrase to the right intent?
- Did its response cause the caller to repeat, simplify, or translate?
Include familiar conversational phrases, short acknowledgements, corrections, negation, and requests to repeat. Test them inside realistic turns instead of as isolated vocabulary. A recognizer may handle a word in a clean sample and still miss it when it appears between English phrases, after an interruption, or beside a customer reference.
Review the audio with the transcript. Text alone cannot show hesitation, repetition, or the moment a caller abandons Twi. Keep the consent record and event trail beside the call so reviewers can connect what was heard, what the agent inferred, and what happened next.
Asenda Talk provides native Twi speech recognition and synthesis fine-tuned in-house, plus configurable agent personas, first messages, voices, and call audit trails. The platform remains in active early access. Its Twi performance should be evaluated against real call patterns, and paid outbound calling remains behind an operator-controlled gate while the live telephony-provider decision is unresolved.
Make language switching a test result
A practical evaluation sheet should give each language switch a reason code: natural code-switching, agent-requested clarification, recognition failure, wrong intent, or caller accommodation. Reviewers can then separate ordinary bilingual speech from moments where the system pushed someone toward English.
Return to Avianca Flight 052 for the operational lesson. The words exchanged during a critical sequence were insufficient because the receiver did not assign them the urgency the speaker intended. More conversation did not repair the underlying interpretation in time.
For a Twi voice agent, the next useful test is specific: find the first phrase that makes a caller switch to English, inspect the preceding turn, and add that phrase in context to the next evaluation set. The correction belongs in recognition and intent testing, not in a report that merely says the call completed.
Comments
No comments yet.