A language switch does not reset the conversation. The caller’s last English sentence carries the intent, subject and emotional context needed to interpret the Twi that follows.
In 1940, Alan Turing and his colleagues at Bletchley Park faced German Enigma messages that looked like meaningless strings. The machine’s settings changed regularly, leaving the codebreakers with too many possible configurations to test blindly. They needed a reliable piece of context.
The words around the unknown words
The codebreakers used “cribs,” educated guesses about phrases likely to appear in a message. A routine report might contain familiar language or follow a known format. That small piece of probable meaning gave the team something against which to test possible Enigma settings.
A crib did not decode the message by itself. It narrowed the search.
Turing’s work on the cryptanalytic Bombe helped automate that search. The machine tested possible settings for contradictions against the expected text. When a candidate survived, the codebreakers could investigate further. Accounts of this work are preserved by the UK’s Bletchley Park Trust and in Turing’s wartime papers held by The National Archives.
The mechanism matters here. A difficult sequence became more interpretable when the system retained a clue about what had come immediately before it.
That is also what happens when a caller says, “I’ve already explained this twice,” then switches to Twi.
The Twi utterance does not arrive in isolation. The English sentence tells us the caller is likely correcting, rejecting or expressing frustration. Discard that sentence at the language boundary, and the system loses part of the evidence it needs.
A switch changes the language, not the intent
Consider a customer calling from Kumasi about a disputed payment. She begins in English:
“I was charged again after I paid.”
Then she continues in Twi.
A speech system can produce an accurate transcription of the Twi words and still misunderstand the turn. Is she repeating the complaint, asking what happens next, rejecting an explanation or telling the agent to stop? The answer may depend on the sentence before the switch, her previous turns and the agent’s last question.
Emotion travels across the boundary too. A short Twi response after “You people keep calling me” should not be treated like a fresh, neutral request. The preceding English establishes resistance. If the Twi includes an opt-out, the system must detect it, record it and prevent another question from exposing information or extending an unwanted call. That concern is explored further in Abena’s Twi opt-out. One more question risks exposing her private information.
This is why code-switch evaluation cannot stop at word error rate. The useful question is whether the agent preserved the caller’s meaning across both languages.
What a voice agent must carry across the switch
A bilingual agent needs a shared conversational state. At minimum, that state should preserve:
- The agent’s previous question and the caller’s last English sentence.
- The active subject, such as a payment, delivery, account or appointment.
- The caller’s apparent intent before and after the switch.
- Negation, correction, consent and opt-out signals.
- The uncertainty attached to the interpretation.
That final point matters. A system should not convert an uncertain reading into a confident action simply because it recognized every word. When the evidence conflicts, the agent can ask a narrow clarification. When the caller has withdrawn consent or asked to stop, clarification must not become an excuse to continue.
Asenda Talk approaches this problem with native Twi speech recognition and synthesis fine-tuned in-house, alongside configurable agent personas, first messages and voices. Vapi orchestrates the assistant runtime. The platform also records call lifecycle events, consent, opt-outs and audit history so teams can inspect what the system heard and what it did next.
Asenda Talk remains in active early access. More African languages are in progress, and feature parity with established voice-agent platforms has not yet been reached. Outbound calling also remains behind an operator-controlled real-money gate while the live telephony-provider decision is unresolved.
Test the boundary, not only each language
A useful evaluation set should include whole conversational turns. Give the system the English lead-in, the Twi continuation and the agent question that prompted both. Then score the resulting action.
Did it keep the same subject? Did it preserve the negative? Did it recognize that a calm explanation had become a complaint? Did it stop when the caller opted out? Could an auditor reconstruct the decision from the stored events?
Testing English and Twi separately misses the failure at the seam. The same applies when reviewing real calls: a transcript segmented by language can hide the evidence that explains the next utterance. What Happens When a Customer Switches Between Twi and English During a Dispute? examines that operational problem in more detail.
At Bletchley Park, the crib supplied a constraint that made an otherwise enormous search tractable. In a bilingual call, the sentence before the switch plays a similar role. Keep it attached to the Twi that follows, test the combined meaning and record the action the agent took.
Comments
No comments yet.