A Twi-English voice agent must retain the caller’s intent, constraints, and consent state across the switch. Changing recognition or synthesis language alone does not establish that the system understood what the caller meant.
In 1999, NASA lost the Mars Climate Orbiter after one part of the mission used imperial units while another expected metric units. The spacecraft had reached Mars, but the data exchanged between teams carried different meanings. NASA’s Mars Climate Orbiter Mishap Investigation Board documented the failure: the values were present, yet their interpretation broke the mission.
A caller code-switching mid-sentence creates a smaller but structurally similar problem. The audio may be clear. The transcript may contain recognisable English and Twi words. Still, the agent has to determine whether the caller is confirming an appointment, changing a number, asking for help, refusing a call, or withdrawing consent.
Meaning must survive the switch
Consider a caller who begins in English, then moves into Twi while clarifying a detail. The agent’s job is to follow the complete instruction as one turn of conversation, rather than treating the Twi phrase as a detached language event.
That requires more than selecting a different voice. Recognition has to capture the words accurately enough for the conversation runtime to interpret them in context. The runtime then needs the right state from the earlier English phrase: who the caller is discussing, what action is underway, and which restrictions already apply.
This matters most when a switch contains a constraint. A caller may start by agreeing to hear more, then use Twi to limit when they can be contacted or to opt out. If the agent hears the language but loses the constraint, the call has failed in the place that matters: the resulting action no longer matches the caller’s meaning.
Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house, alongside an assistant runtime orchestrated through Vapi. That is the work under evaluation today. More African languages are in progress, and the platform remains in early access.
The transcript is evidence, not the whole answer
A transcript can help a team investigate a call. It cannot, by itself, prove that the agent took the correct next action.
For voice teams, the operational question is straightforward: after a caller switches languages, can you reconstruct what the system heard, what it inferred, and what it did? That is why call-truth tracking and an audit trail belong beside language performance. If a caller says no, asks to be removed, or corrects a detail, the record needs to connect that moment to the action that followed.
Asenda Talk includes consent, opt-out, and audit-trail handling for every call, plus telephony lifecycle webhooks intended to track the call through its stages. Those controls give a support or campaign team something concrete to inspect when a conversation goes wrong.
The useful test is not “did the agent speak Twi?” Ask whether the record can show that it retained the caller’s restriction after the switch. Keeping caller constraints across language switches is where this becomes especially important.
Test the moments where language carries the decision
Teams evaluating a Twi-English agent should build test calls around decision points, not polished greetings. Include a caller who changes a date halfway through a sentence. Include one who repeats a phrase more slowly. Include one who starts in English and gives the decisive correction in Twi.
Then review three things:
- Did the system recognise the switch accurately enough to preserve the instruction?
- Did the agent retain context from the earlier part of the turn?
- Did the resulting call record show the correct outcome and any consent change?
These tests are more useful than a general demo because they expose the handoff between speech recognition, conversation state, and downstream call handling. They also reveal where a platform needs more evaluation before it belongs in a live workflow.
Outbound calling through Asenda Talk is not live by default. It remains gated behind an explicit telephony-provider decision, and the platform’s real-money per-minute billing is controlled by an operator gate. That constraint is deliberate: a good conversation demo does not remove the need for clear approval, auditable records, and a provider decision before a campaign line goes live.
Build for interpretation before scale
The Mars Climate Orbiter did not fail because its teams lacked numbers. It failed because matching values were interpreted through incompatible systems. A bilingual voice workflow faces its own version of that risk when language changes but intent does not travel with it.
Before adding more callers or more call volume, test the exact moments where Twi changes the meaning of an English conversation. Keep the transcript, the inferred intent, the final action, and the consent record close enough to inspect together. That gives your team a way to find the mismatch while it is still a test case, rather than after a caller has been misunderstood.
Comments
No comments yet.