African voice AI should be evaluated across four separate questions: speech-model quality, code-switching behaviour, telephony readiness, and deployability. A language count can signal coverage, but it does not show whether a caller will be understood, handled safely, or connected to a live phone line.
At 4:40 p.m. in Kumasi, Adjoa is holding a paper list of customers who asked her cooperative to call them back about their applications. The agent she is testing opens in Twi, but the caller answers with a sentence that shifts into English for the product name, then back into Twi to explain a concern. On the second turn, the agent treats the concern as a new application instead of a request for clarification.
The day’s call list is waiting. If the team mistakes these intents at scale, people who expected help could receive the wrong follow-up, and the cooperative could lose trust before the campaign has properly begun. A provider’s long language list cannot answer the question Adjoa needs answered: did the system preserve the caller’s meaning through the switch?
Evaluate the speech model, not the label beside the language
“Supports Twi” can describe very different systems. It may mean a general multilingual model has some Twi coverage. It may mean speech passes through a third-party voice API. Or it may mean the team has built and evaluated Twi recognition and synthesis directly.
Those differences affect the work before a deployment. Test real phrases your callers use, including names, local place references, product terms, confirmations, corrections, and speech spoken more slowly after the agent misses it. Listen to the audio as well as reading transcripts. A transcript can look close enough while the caller hears an unnatural pronunciation or an awkward pause that changes the tone of the exchange.
Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house. That is the current product focus, with more languages in progress. It gives teams a specific basis for evaluation: recognition and synthesis for Twi, rather than a broad claim about African-language coverage.
The right question is not, “How many languages are on the page?” Ask, “Can this system handle the language my callers actually speak in this workflow?”
Test code-switching as an intent problem
Code-switching is not a cosmetic detail. A caller can begin in Twi, use English for an account type or product name, then return to Twi when describing what they need. The agent must carry the same constraint across those turns.
For Adjoa’s cooperative, the important test is not whether the agent recognizes each word in isolation. It is whether it keeps track of the caller’s request after the switch. Did the caller ask for a callback? Did they say they already applied? Did they withdraw consent? Did they correct the agent?
Build a test set around those moments. Include a greeting repeated more slowly, a correction after an incorrect assumption, and a request that changes language mid-sentence. Then define what a safe outcome looks like. In some flows, the correct outcome is a clear confirmation. In others, it is handing the call to a person or ending without another marketing prompt.
This is also where call records matter. If a team cannot inspect what the agent heard, said, and decided, it will struggle to investigate an error after the fact. Twi English Voice Agents: Keeping Caller Constraints Across Language Switches explores that continuity problem in more detail.
Separate telephony capability from a voice demo
A convincing browser demo does not prove that a voice agent is ready to call real phone numbers. Telephony introduces provider decisions, call status events, consent handling, opt-outs, retries, billing, and the evidence needed when a customer asks why they were contacted.
Asenda Talk has a telephony lifecycle webhook pipeline for call-truth tracking, plus consent, opt-out, and audit-trail handling for every call. Its runtime is orchestrated through Vapi, and the platform includes metered per-minute billing with an operator-controlled real-money gate.
Those are useful foundations, but they are not a claim that outbound calling is already live. Asenda Talk is in active early access, and outbound calling remains gated by an explicit telephony-provider decision that has not yet been made live. A team evaluating any platform should ask which call paths are available now, which depend on provider approval or configuration, and what evidence will exist after a disputed call.
For a consent-sensitive workflow, traceability may matter as much as the voice itself. Support manager first hour: How One Traceable Call Record Establishes Consent shows the practical questions that record should answer.
Judge deployability by the controls around the conversation
Deployability means a team can configure an agent, protect credentials, control spending, and investigate what happened without improvising around the platform. It should be evaluated separately from language quality and telephony reach.
Adjoa returns to the test later with a smaller call set and clearer scenarios. The agent’s persona, first message, and voice are configured for the flow. When a caller changes language to correct an assumption, the team checks whether the intended constraint survives before expanding testing. They also confirm how secrets are stored, how the call trail is reviewed, and who can open the real-money gate.
That is a more useful readiness test than comparing language counts. Choose the speech model by listening and measuring comprehension. Choose code-switching behaviour by testing meaning across turns. Choose telephony by verifying the live path and its records. Choose deployability by checking the controls your operators will actually use when something goes wrong.
Comments
No comments yet.