Choosing Vapi answers how a voice agent is orchestrated at runtime. It does not decide who recognizes Twi, generates Twi speech, connects the call, or records what actually happened.
At 4:40 p.m. in an Accra meeting room, Kojo, a fictional product lead with a cooling mug of tea beside his laptop, reached the final box in an architecture diagram. The team had already written “Vapi” across the centre and assumed the main decision was complete. Then a developer played a test call in which the caller switched from English to Twi halfway through a consent statement.
The transcript lost the meaning of the reply. If the team approved the design that evening, the first pilot could record consent that the speech layer had never understood. The launch date on the whiteboard suddenly looked doubtful.
One voice call contains three separate systems
The review changed when Kojo erased the single “voice platform” box and drew three.
The orchestration layer controls the assistant’s runtime behaviour. It manages prompts, tool calls, turn-taking and the flow of a conversation. Asenda Talk uses Vapi for this layer.
The speech layer handles recognition and synthesis. It turns a caller’s audio into text, then turns the agent’s response back into audio. This is where language quality lives. A system can orchestrate a call correctly while mishearing the Twi words that determine consent, an appointment date or the amount being discussed.
The telephony layer connects the call to a phone network. It also produces lifecycle events such as ringing, answered, completed and failed. Those events need reconciliation because a provider’s “completed” status may describe a technical connection rather than a successful conversation. What Really Happened When the Bilingual Call Was Marked Complete? examines that gap in more detail.
Treating these layers as one purchasing decision hides ownership. A team may choose an orchestration platform and still have no tested answer for Twi recognition, Twi synthesis or live call delivery.
Put an owner beside every language boundary
Kojo’s team returned to the recording. This time, they traced the audio through the proposed system.
Who receives the caller’s raw speech? Which component identifies the words? What happens when the caller moves from Twi to English and back again? Which system produces the voice response? Where can the team inspect the transcript, audio and call events when the result is disputed?
Those questions turned the diagram into an ownership map.
For Asenda Talk, native Twi recognition and synthesis are fine-tuned in-house. They are separate from Vapi’s role in orchestrating the assistant runtime. That distinction matters because changing an orchestration component should not quietly replace the language models being evaluated.
The same discipline applies to telephony. Asenda Talk has a webhook pipeline for call lifecycle events, call-truth tracking, metered per-minute billing and an operator-controlled real-money gate. A live outbound telephony provider has not yet been selected, so outbound calling remains gated in early access. The diagram must show that unresolved decision plainly.
A useful review labels each connection with four things: the data crossing it, the system responsible, the evidence retained and the failure behaviour. “Audio to speech recognizer” is more useful than an arrow labelled “AI.” “Provider event reconciled against internal call state” is more useful than “call complete.”
Test the handoffs where meaning can break
Architecture diagrams often look clean because they show components, not contested moments. The harder test is to follow one sentence through every boundary.
Use a phrase that could change the outcome of the call. A caller withdraws consent in Twi. A patient corrects a date after switching to English. A customer disputes a charge using a phrase the agent has not heard before.
Then inspect what each layer contributes:
- The speech layer must preserve the caller’s meaning.
- The orchestration layer must choose the correct response or action.
- The telephony layer must report the call state accurately.
- The audit trail must let a reviewer reconstruct the sequence.
- The billing gate must prevent unapproved real-money calling.
Language switching deserves its own test set because a smooth voice can conceal a weak transcript. Twi Voice Agent Testing: What a Caller’s Switch to English Reveals shows why the switch itself is valuable evidence.
Consent and opt-out handling should receive the same treatment. Store what the caller said, how it was interpreted, what action followed and which system produced each record. A green completion badge cannot carry that burden alone.
Leave unresolved decisions visible
With the revised diagram on the screen, Kojo could point to every owned boundary. Vapi controlled runtime orchestration. Asenda’s in-house models owned Twi recognition and synthesis. The lifecycle pipeline owned call-state reconciliation. Consent, opt-out and audit records followed each call. Telephony provider selection remained open, and the real-money gate stayed closed.
The pilot had not become magically ready. The team had found something more useful: the exact decision preventing readiness, before a live caller paid the price for the ambiguity.
At the next review, keep the three boxes separate. Play one bilingual call through them. If anyone answers “the platform handles that,” ask them to name the component, the evidence it leaves behind and what happens when it gets the sentence wrong.
Comments
No comments yet.