A bilingual voice agent must preserve the caller’s intent when Twi and English appear in the same sentence. Translating each fragment separately can produce a correct transcript while sending the conversation down the wrong path.
Consider an illustrative early-access review in Accra. At 4:40 p.m., Efua, a support lead with a cold cup of tea beside her laptop, replayed a test call before approving the next evaluation round.
The caller began in Twi, then ended with an English account number. The transcript looked clean. The agent’s response did not. It treated the account number as a new topic and repeated the opening question, leaving the caller’s request unresolved.
Efua had one review slot left that afternoon. If the next run failed, the bilingual flow would remain blocked rather than move into controlled testing.
One sentence can carry one intent across two languages
A script-translation model often assumes that each language segment can be handled independently. It detects Twi, processes that portion, switches to English, then processes the account number.
That sequence can preserve the words and lose the point.
The caller may be saying, in effect, “Please check the balance for account 4821.” The request lives in the Twi portion. The object of that request arrives in English. Splitting them turns a single instruction into two incomplete fragments.
This is common in natural bilingual speech. A caller may explain a problem in Twi and use English for an identifier, product name, date, department or technical term. They have not started a second conversation. They are completing the first thought with the vocabulary that feels most natural.
The system therefore needs more than language detection. It needs conversational state that survives the switch: what the caller wants, which detail is missing, and how the latest words complete that request.
That is also why a correct transcript can still produce the wrong action. The recognition layer may capture every word while the orchestration layer attaches a detail to the wrong field or intent. Bilingual Voice Agents: Why a Correct Transcript Can Still Trigger the Wrong Action examines that failure more closely.
Test the switch where the meaning depends on both halves
Efua changed the evaluation prompt. She stopped testing separate Twi and English turns and created one mixed-language utterance in which neither half made sense alone.
The first half established the task. The second supplied the account number. The expected result required the agent to join them before responding.
That test exposed three distinct questions:
- Did speech recognition preserve the Twi request accurately?
- Did the system retain that request when the caller switched languages?
- Did it treat the English number as the missing detail rather than a new intent?
These checks should remain separate. If the transcript is wrong, investigate recognition. If the transcript is right but the response is wrong, inspect intent tracking and tool arguments. If both are right but the call record is incomplete, inspect the telephony lifecycle and audit trail.
This separation matters because “the bilingual call failed” is too broad to guide a fix. A team needs to know whether the failure occurred in speech recognition, conversation state, action selection or call reporting.
Asenda Talk’s native Twi speech recognition and synthesis are fine-tuned in-house. Vapi orchestrates the assistant runtime. That architecture gives the team a clear place to evaluate Twi speech quality while also examining how the assistant handles intent across a switch.
The platform is in active early access. More African languages are in progress, and feature parity with established voice-agent platforms remains ongoing work.
High-risk details need confirmation before action
An account number raises the stakes. Even when recognition and intent tracking agree, the agent should confirm sensitive or consequential details before changing a record, disclosing information or triggering another action.
A useful confirmation preserves the whole request: “You want me to check account 4821. Is that correct?” It should not confirm the digits while forgetting what the caller asked the system to do with them.
This pattern also applies when a caller corrects an identifier midway through a sentence. The latest confirmed value must replace the earlier one without erasing the underlying intent. What Happens When an AI Service Misses a Twi Correction Mid-Call? covers the operational risk of missing that change.
Consent, opt-out handling and an audit trail belong in the same evaluation. A fluent response cannot compensate for an unrecorded withdrawal or an action that cannot be traced after the call.
Keep the deployment gate tied to evidence
On Efua’s final run, the system held the Twi request in context, attached the English account number to it, and asked for confirmation. She marked that case as passed.
She did not mark the agent ready for live outbound calling.
Asenda Talk includes telephony lifecycle webhooks, call-truth tracking and metered per-minute billing with an operator-controlled real-money gate. Outbound calling still depends on an explicit telephony-provider decision that has not been made live. Until that decision and the relevant evidence are in place, the gate should remain closed.
For the next review, Efua added more mixed-language cases: a Twi request ending with an English date, a correction delivered after the language switch, and an opt-out expressed halfway through a longer explanation.
The cold tea stayed beside her laptop. The difference was visible in the test record: one intent, one confirmed account number, and no live call approved beyond the available evidence.
Comments
No comments yet.