An English-only voice agent can mishear a Twi greeting before the conversation has properly begun, and callers often respond by repeating the same Twi words more slowly. That instinct matters because a language switch can feel like giving up the natural way they opened the call, rather than fixing a technical problem.
At 8:12 on a humid Monday morning in Kumasi, Abena stands beside her kitchen counter with a school lunchbox still open in front of her. She has called to confirm an appointment. The agent answers in English, and Abena opens in Twi.
The agent asks her to repeat herself.
She does. This time she separates the words, rounds each vowel, and speaks louder than she normally would. The agent offers the same prompt in English.
Abena knows English. She uses it at work. But switching languages is not automatic here. She has begun the call in Twi because that is how she wants to explain the appointment and ask a follow-up question. The important detail may come after the greeting, when she needs to describe a constraint in the way it makes sense to her.
A second failed attempt puts the appointment at risk. If the agent records the wrong date or ends the call without a clear confirmation, Abena may arrive at the wrong time and have to start again. She pauses, looking at the lunchbox lid in her hand, deciding whether to force the rest of the conversation into English.
That pause is the first repeat.
A repeat is a signal about recognition, not willingness
An agent that hears a Twi greeting as unclear speech can treat the repeat as a prompt to push the caller toward English. The caller may mean something else entirely: “I said what I meant. Please hear it.”
People naturally adapt when they think a listener missed the sound of a word. They slow down. They repeat. They change emphasis. They do not necessarily translate.
For a support desk or campaign team, that distinction changes how an early turn should be handled. A repeated Twi phrase should remain available to the speech system as Twi. It should not become evidence that the caller has failed to comply with an English-first conversation.
The cost is larger than an awkward greeting. A forced language change can shorten an answer, remove a correction, or make a caller accept an incomplete outcome because explaining it again feels tiring. The transcript may show a completed call while the person leaves without the answer they needed.
That risk becomes sharper when Twi and English appear in the same call. A caller may use English for a date, a name, or a product term, then return to Twi for the part that carries intent. Twi and English code-switching: Why voice agents must preserve context explores why the change in language must not break the meaning of the conversation.
The greeting sets the caller’s terms for the rest of the call
The first seconds of a call tell the caller what kind of interaction they are in. If an agent recognizes the greeting and responds appropriately, the caller can continue without first negotiating language access.
If it does not, the caller starts doing extra work. They may pronounce words unnaturally, simplify their request, or switch languages before they are ready. Those adaptations can make an English-only system appear functional in a call log while hiding the friction that produced the outcome.
Abena eventually tries English. She gives the appointment details in short pieces and leaves out the follow-up question she wanted to ask. The agent confirms a date. On paper, the call has reached an end state. For Abena, it has created another task: finding a person later who can answer the question in Twi.
A useful voice agent should recognize that the caller’s first language choice contains information. It says how they expect to speak, what level of explanation they may need, and what language should stay available when the conversation becomes more detailed.
Build for the turn after the first misunderstanding
Asenda Talk is built around native Twi speech recognition and synthesis, fine-tuned in-house. The goal is to give builders a foundation for Twi and English conversation that does not begin by treating Twi as an exception to an English call.
That foundation needs careful agent design. Configure the agent’s first message and persona so callers know Twi is welcome. Test greetings, corrections, and short repeated phrases, not only long scripted answers. Review where the transcript and call outcome diverge. A call that reaches its final webhook event may still contain an unresolved question or a language-driven misunderstanding.
For calls that require consent or an opt-out, the stakes are higher. The record must preserve what happened, including the caller’s language and the point at which they asked to stop. Asenda Talk includes consent, opt-out, and audit-trail capabilities for every call, along with telephony lifecycle webhook tracking. Those controls support review; they do not replace careful language evaluation.
Outbound calling is still gated behind an explicit telephony-provider decision that is not live. Teams evaluating Asenda Talk today can build and configure agents, assess Twi speech behavior, and prepare their call handling before treating outbound deployment as available.
Test the phrase callers say twice
Abena’s call should have taken a different turn on the first repeat. The agent should have continued to listen for Twi, then let her complete the appointment question in the language she chose. Later that morning, she would have one confirmed date, one answered question, and no reason to call again.
Before putting a voice flow in front of callers, collect the phrases people use to open, correct, hesitate, refuse, and clarify. Have Twi speakers say them at normal speed, then repeat them the way they would after a machine misunderstands them. Review both the recognition result and the agent’s next action.
The test is simple: when a caller slows down in Twi, does the agent hear a request to be understood, or an invitation to abandon Twi?
Comments
No comments yet.