Jim Lovell’s Canisters Do Not Fit. Carbon Dioxide Keeps Accumulating.

A voice agent has failed when a caller understands the welcome but cannot follow the next instruction. Bilingual testing must cover the entire task, especially the language changes that determine whether the caller can act.

In April 1970, Apollo 13’s crew faced a compatibility problem inside the spacecraft. After an oxygen tank exploded, astronauts Jim Lovell, Jack Swigert and Fred Haise moved into the lunar module while NASA’s Mission Control worked to keep them alive. Carbon dioxide began accumulating. The command module carried square lithium hydroxide canisters, while the lunar module used round openings.

Both systems worked as designed. They could not work together when the mission depended on them.

Engineers in Houston had to devise an adapter using materials available aboard the spacecraft. NASA documented the improvised carbon dioxide scrubber procedure, and Lovell later described the mission in Lost Moon, written with Jeffrey Kluger. Before the adapter was built and tested, compatibility was an unresolved survival problem.

The stakes of a voice-agent launch are smaller. The mechanism is familiar: one working component can hand a person into a step they cannot complete.

A successful greeting can hide a broken task

Imagine the first call on launch morning. The agent gives its welcome in English. The caller understands why the business is calling and waits for the next instruction.

That instruction arrives in English too. It asks the caller to choose an option, confirm a detail or explain what happened. The caller hesitates. When the same step is given in Twi, the caller responds.

The greeting passed. The task had already failed.

A team reviewing only the opening might mark the agent as bilingual because it can speak Twi or because the welcome sounds natural. That test says little about whether the caller can finish the job. The meaningful test starts when the conversation asks for action.

Can the caller correct a name in Twi? Can they withdraw consent after beginning in English? Can they explain a support problem in one language and confirm the next step in another? Those transitions decide whether the agent records the right outcome.

This is also why a correct transcript can still produce the wrong operational result, as discussed in Bilingual Voice Agents: Why a Correct Transcript Can Still Trigger the Wrong Action.

Test the instruction, response and recorded outcome

A useful bilingual test follows one complete task from the first word to the final system state.

Start with an English greeting, then give the action instruction in Twi. Reverse the languages on the next run. Let the caller begin an answer in English and finish it in Twi. Include corrections, interruptions and a direct request to stop the call.

Then inspect more than the audio.

Check what the speech system recognized. Check how the assistant interpreted it. Check what action followed. Finally, inspect the call record, consent state, opt-out status and audit trail. A natural-sounding exchange can still leave the wrong value behind.

Asenda Talk is built for this layer of the problem. Its Twi speech recognition and synthesis are fine-tuned in-house, rather than passed through a generic third-party speech wrapper. Teams can configure an agent’s persona, first message and voice, while the telephony lifecycle pipeline records what happened during the call.

That does not make every bilingual path correct by default. Twi recognition, reasoning, calling orchestration and downstream actions must be evaluated together. What Happens When a Caller Finishes an English Thought in Twi? examines one version of that handoff.

Treat language transitions as launch gates

Before a campaign or support line goes live, write down the moments where misunderstanding would change the result. Account changes, payment discussions, appointment confirmations, consent and opt-out requests deserve direct tests in every supported language path.

Each test should have a deterministic expected outcome. If the caller says no in Twi, the record should show no. If the caller corrects a detail after switching languages, the final value should reflect the correction. If recognition confidence or agent behavior is unclear, route the case for human review instead of treating a completed call as proof of success.

Asenda Talk remains in active early access. Twi capability is built and available for evaluation, while more African languages are in progress. Outbound calling also remains gated behind an explicit telephony-provider decision that has not been made live. A launch plan should preserve those distinctions and report blocked gates plainly.

Apollo 13’s scrubber problem was solved only after engineers tested the whole connection between the available canister, the adapter and the lunar module opening. Voice teams need the same discipline at lower stakes: test the point where one working part must connect to the next.

For the next review, choose one high-consequence caller task. Run it English to Twi, Twi to English and mixed within one response. Do not approve the path until the caller’s words, the agent’s action and the audit record agree.