A resolved call needs more than a confident intent label. A support manager should compare the caller’s audio, the Twi transcript, the assistant’s inferred intent, and the call event record before deciding what happened.
Consider an illustrative case. At 8:07am in Accra, support manager Adwoa Mensah has one earbud in and a mug of tea cooling beside her keyboard. The dashboard says a caller’s billing question was resolved. The transcript appears to support that result.
Then Adwoa plays the audio.
Two versions of the same call
The assistant inferred that the caller accepted an explanation about a second charge. Read quickly, the transcript suggests agreement too. A short Twi response has been rendered in English as confirmation, followed by a polite closing.
The caller’s voice tells a different story.
Adwoa hears hesitation before the response. The caller repeats the reference to the second charge, stresses one word, then switches briefly into English. The assistant continues toward closure. By the end, the call record presents a tidy outcome: question answered, caller acknowledged, conversation ended.
Adwoa cannot approve that outcome yet. If the caller was correcting the assistant rather than agreeing with it, closing the case could leave a disputed charge unexamined. The next person reviewing the account might see “resolved” and move on.
For one uncomfortable minute, both versions remain plausible.
This is where bilingual voice support becomes an evidence problem. Audio carries emphasis, timing, interruption, and uncertainty. A transcript turns those signals into text. An intent model compresses the text again into a category. Each step can be useful, but each step can also remove context.
The issue becomes sharper when the hardest part of the conversation switches languages. What happens when a support call switches to Twi is rarely captured by an English-only summary.
The transcript is evidence, not the verdict
Adwoa slows the recording and compares it with the Twi transcript line by line. Asenda Talk’s Twi speech recognition and synthesis are fine-tuned in-house. That gives the team direct responsibility for evaluating how the speech layer handles Twi, rather than treating a third-party transcript as an unquestionable result.
She finds the crucial difference. The assistant interpreted the caller’s phrase as acceptance of the explanation. In context, the caller appears to be distinguishing the second charge from the first one already discussed.
That distinction changes the case.
Adwoa does not need the system to pretend certainty. She needs it to preserve enough evidence for a human to challenge the label. The original audio, transcript, inferred intent, and call lifecycle events should remain connected to the same call. If one source conflicts with another, the conflict belongs in the review process.
The lesson also applies beyond billing. A short Twi response could signal consent, refusal, correction, confusion, or a request to stop. Flattening those possibilities into a neat English category can create operational risk. A customer’s Twi evidence can lose one detail and change a card dispute.
Call truth needs a chain of evidence
Asenda Talk includes a telephony lifecycle webhook pipeline with call-truth tracking. It records what the calling system reports across the call lifecycle, giving reviewers an event trail to compare with the conversational record.
That matters because “resolved” can refer to several different facts. The assistant may have completed its planned dialogue. The call may have ended normally. The intent model may have assigned a resolution label. None of those facts proves that the caller’s underlying issue was settled.
A defensible review asks separate questions:
- What can be heard in the original audio?
- What did the Twi speech recognizer transcribe?
- What intent did the assistant infer from that transcript?
- Did the caller consent to continue, or attempt to opt out?
- Which lifecycle events occurred before the call ended?
- What evidence supports the final case status?
This separation reduces the temptation to let one green status badge settle every question. It also gives language and support teams something specific to evaluate. They can inspect the phrase, the inferred intent, and the event sequence instead of debating a vague complaint that “the AI got it wrong.”
Consent and opt-out records deserve the same treatment. A completed call does not erase a refusal, and a polished closing does not prove consent remained valid throughout the conversation. Asenda Talk keeps an audit trail for each call so those decisions can be reviewed.
Mark uncertainty before closing the case
With the audio and Twi transcript aligned, Adwoa changes the illustrative case from resolved to needs review. She leaves a short note identifying the disputed phrase and the assistant’s conflicting interpretation. The case stays open long enough for a human follow-up.
At 8:19am, the tea is still untouched. The difference is that the next reviewer will see the uncertainty instead of inheriting a false conclusion.
Asenda Talk is in active early access. Teams can create and configure voice agents, including persona, first message, and voice, while the platform’s native Twi speech layer and call records are evaluated. Vapi orchestrates the assistant runtime. Paid outbound calling remains behind an operator-controlled real-money gate, pending an explicit telephony-provider decision.
That stage calls for disciplined review. Before automating closure, define which intent labels require audio checks, what level of transcript uncertainty triggers human review, and which consent or opt-out events must block further action.
The next 8:07am call may still produce two versions of the truth. The system should make sure Adwoa can find both before someone closes the case.
Comments
No comments yet.