Bilingual Voice Agent Testing: Why an Account Match Cannot Prove Understanding

A matched account number can create a false signal of success when the voice agent fails to understand the customer’s explanation. In a bilingual call, identity verification and problem resolution must be measured separately.

In 1999, NASA’s Mars Climate Orbiter passed behind Mars and failed to re-establish communication. The mission had reached its destination, but one interface had concealed a critical mismatch: Lockheed Martin supplied thruster data in pound-force seconds while navigation software at NASA’s Jet Propulsion Laboratory expected newton-seconds.

The values moved through the system. The calculations ran. Progress appeared real until the spacecraft was lost.

A successful check can hide a failed conversation

Picture a composite morning call with a customer in Kumasi.

An agent asks for an account number in English. The customer reads it aloud. The system recognizes the digits, finds the record and marks identity verification complete.

Then the customer switches to Twi to explain the actual problem. A payment appears twice. One entry may be pending, or the customer may be disputing both. The distinction sits inside a longer explanation, with English financial terms embedded in Twi.

The agent misses that distinction. It catches a familiar word, selects a likely intent and responds with a standard explanation about pending transactions. The account matched, so the workflow looks healthy. The customer repeats the point in Twi, receives another answer to the wrong problem and ends the call.

Several events may still appear positive in an operational dashboard:

• The call connected.
• The account number matched.
• The agent selected an intent.
• A response played.
• The call ended without a technical error.

None proves that the system understood why the customer called.

This is the same structural failure exposed by the Mars Climate Orbiter investigation. Each component could appear to perform its assigned task while the meaning crossing the boundary was wrong. Arthur Stephenson chaired the mishap investigation board, whose 1999 Phase I report documented the unit mismatch and the failures that allowed it to pass undetected.

For voice AI, the boundary is often linguistic rather than numerical. The transcript may contain plausible words. The workflow may advance. Yet the customer’s meaning can disappear when the conversation moves from English account details into Twi explanation.

Track understanding after the language switch

A bilingual voice agent needs more than a call-completed status. Teams should inspect what happens at the point where the customer changes language, especially when that switch introduces the complaint, correction, consent decision or payment dispute.

Start by separating lifecycle facts from conversation outcomes. A telephony webhook can establish that a call connected, continued and ended. It cannot establish that the agent understood the customer.

Then test the difficult span directly. Compare the audio, Twi transcript, detected intent, agent response and final disposition. If the agent records “payment pending” after the customer disputed a duplicate charge, the failure must remain visible even when every telephony event arrived correctly.

This is why Twi voice agent testing should examine what a caller’s switch to English reveals, and why the reverse switch deserves the same scrutiny. Language choice can be evidence. A customer may change languages because the first explanation failed, because a precise term comes more naturally in another language, or because the issue becomes harder to describe.

Those possibilities should become test cases, not assumptions about the caller.

Design completion around the customer’s problem

For the composite call, a defensible completion record would need to show more than the account match. It should preserve which issue the customer described, what the agent understood, how the agent responded and why the final status was assigned.

A practical review set can include calls where:

• The account identifier is spoken in English and the problem in Twi.
• The customer corrects the agent without changing the account details.
• Financial terms remain in English inside a Twi sentence.
• The agent asks for confirmation before assigning a disposition.
• The customer abandons after repeating the same explanation.

Consent and opt-out events require the same discipline. If the customer says to stop calling during a misunderstood Twi passage, a successful account lookup cannot override that instruction. The audit trail must make the sequence reviewable. A call review should prove what happened after a caller says “stop calling”, not rely on a green completion marker.

Asenda Talk is being built for this kind of inspection. Its current foundation includes native Twi speech recognition and synthesis fine-tuned in-house, configurable voice agents, telephony lifecycle tracking, consent and opt-out records, and per-call audit trails. It remains in active early access. Outbound calling is still behind an operator-controlled real-money gate while the live telephony-provider decision remains open.

That caveat matters because reliable evaluation should precede scale. More calls only multiply whatever the completion logic rewards.

Test the boundary before trusting the status

The Mars Climate Orbiter report did not treat a received value as proof that both sides agreed on its meaning. Voice teams should apply the same rule.

Build a small bilingual evaluation set from approved, consented test calls. Mark the account-verification span separately from the problem-description span. Have a fluent Twi reviewer judge whether the transcript preserves the complaint, then compare that judgment with the agent’s intent and disposition.

Finally, make abandonment visible as an outcome. A customer who leaves after two misunderstood explanations did not complete a support journey, even if the system found the right account in the first ten seconds.

The most useful next test is simple: take every call marked complete after a language switch and verify the problem independently. The boundary deserves inspection before the green checkmark deserves trust.