A compliance lead can reconstruct a disputed voice-agent call only when the evidence connects four separate records: what the agent produced, what speech recognition captured, what the system retained, and what action followed. A transcript alone cannot establish that chain.
At 4:47pm in Accra, Esi receives a complaint from a caller who says a voice agent ignored his Twi opt-out and continued discussing an account issue. Esi is a composite compliance lead, the sort who keeps tea beside her keyboard until it goes cold. Her manager needs an answer before the team closes for the evening.
The complaint leaves two bad possibilities on the table. The agent may have continued after a clear refusal, exposing a consent failure. Or the caller’s Twi response may have been misunderstood, sending Esi toward the wrong conclusion with a polished English transcript as her only evidence.
She opens the call record. “Completed” tells her almost nothing.
A call record needs four connected layers
Esi starts with the agent output: the exact text the assistant generated and the audio synthesis request associated with it. This establishes what the system intended to say. If the caller heard different wording because audio was clipped, delayed or interrupted, the generated text cannot settle the complaint by itself.
Next comes the recognition layer. What audio reached speech recognition? What words did the recognizer return, in which language, and with what timing? In a Twi and English conversation, normalized text can hide the phrase that changed the caller’s meaning. A clean transcript can also conceal an earlier opt-out.
The third layer is the retained record. This may include the transcript, consent event, opt-out event, call identifiers and timestamps. Each item must point to the same call. A transcript copied into a support ticket without its original call ID creates a second story that may drift from the first.
The fourth layer is action. Did the system mark the caller as opted out? Did it attempt another turn? Did the telephony provider end the call, report an error or keep the connection open? Call status must come from lifecycle events, not from an assistant message claiming that the call ended.
These layers create a sequence Esi can test:
- The agent generated a consent prompt.
- Speech recognition captured the caller’s response.
- The system recorded the consent or opt-out state.
- The runtime acted on that state.
- Telephony lifecycle events established what happened to the call.
If one link is absent, the finding should say so plainly.
Timestamps turn separate logs into evidence
At 5:06pm, Esi finds the decisive gap. The transcript contains an English rendering of the caller’s response, but the opt-out audit event appears several seconds later. Between them sits another generated agent turn.
Now the outcome is genuinely uncertain. If the caller opted out before that turn, the agent continued when it should have stopped. If recognition produced the opt-out only afterward, the fault may sit earlier in the speech pipeline. Without aligned timestamps and stable identifiers, Esi cannot tell which sequence occurred.
This is why call-truth tracking matters. Asenda Talk has a telephony lifecycle webhook pipeline designed to capture provider-reported call events and connect them to the platform’s call state. It also includes consent, opt-out and audit records for each call. Native Twi speech recognition and synthesis are fine-tuned in-house, which gives the team direct responsibility for evaluating how Twi audio becomes text and speech rather than treating language behavior as an opaque third-party wrapper.
That architecture creates the pieces needed for investigation. It does not make every investigation complete by default. Retention rules, recording policy, provider payloads and the visibility of each event still determine what a compliance lead can prove.
Asenda Talk remains in active early access. Vapi orchestrates the assistant runtime, while the live telephony-provider decision for outbound calling has not been finalized. Outbound calls remain behind an operator-controlled real-money gate. Any evidence plan must therefore be tested again against the provider chosen for live calling, including its event ordering, identifiers and failure states.
Preserve disagreement instead of cleaning it away
A compliance record becomes weaker when conflicting evidence is silently reconciled. If the recognized Twi differs from a later English summary, keep both. If the assistant generated a closing line but the provider reports that the call remained connected, retain the disagreement. If an opt-out arrives after another agent turn, do not reorder the events to match the expected flow.
The audit trail should also record relevant configuration. Which agent persona, first message and voice were active? Which environment handled the call? Which configuration version applied at that moment? Secrets should remain write-only and masked. Their values do not belong in an investigation export, but the record can identify which approved secret location and environment were used.
This separation protects the evidence from convenient explanations added later. The same principle applies to consent: a dial list cannot prove that a caller granted permission during a call.
End with a finding the evidence can support
By 5:31pm, Esi has a bounded answer. She can show the generated prompt, the recognized response, the timestamped opt-out event, the next runtime turn and the provider lifecycle events. Where the sequence remains ambiguous, she labels the gap instead of guessing.
Her finding does more than classify one complaint. It produces a test case for the Twi recognizer, a check for opt-out timing and a requirement for the eventual telephony provider. Before the next live call, the team can replay the sequence in a controlled evaluation and verify that an opt-out blocks another turn.
Esi closes the complaint record with the evidence attached and the unresolved point named. The cold tea is still beside her keyboard. This time, “completed” has a traceable meaning.
Comments
No comments yet.