An “African voice agent” can still depend on English-first speech recognition and synthesis underneath. Technical buyers should test the speech layer directly with the language, accents, code-switching, and call openings their customers use before they judge the orchestration layer.
At 3:40 p.m., Kojo was alone in a meeting room in Accra with a cooling cup of sobolo beside his laptop. He had been asked to assess a voice-agent platform for a support desk whose callers often moved between Twi and English in the same explanation.
The demo had looked convincing. An agent could be configured with a persona, a first message, and a voice. It could pass a call through an assistant runtime. The dashboard showed familiar call states.
Then Kojo opened the technical documentation and traced the audio path.
The platform described itself as built for African conversations, but its language layer centred English. Twi appeared as a prompt instruction, a translation step, or an external integration. There was no clear account of speech recognition trained for Twi, no synthesis model trained to produce Twi naturally, and no evaluation evidence for the speech his callers would actually hear.
His deadline was the next morning. If he approved the wrong architecture, the support desk could launch a service line that sounded competent in a scripted English demo and broke down when a caller explained their problem in Twi. The failure would land with the caller, not with the diagram in the vendor deck.
Orchestration and speech are separate technical decisions
Voice-agent platforms do important work. They manage prompts, tools, call state, webhooks, routing, analytics, and the assistant logic that decides what to say next. Vapi, Retell, Bland AI, and similar platforms have made that layer more accessible.
But orchestration does not determine whether a caller’s speech is understood correctly. It does not make a synthesized voice sound natural in Twi. Those outcomes depend on the speech recognition and speech synthesis systems handling the audio.
That distinction matters when a product uses African-language positioning. A platform may be able to place a language name in a prompt while the underlying recognizer was primarily built for English. It may generate an answer whose text is plausible, then produce speech that has weak pronunciation, unnatural rhythm, or poor handling of common language switching.
For Kojo, the useful question changed from “Can this platform run a voice agent?” to “Which component receives the caller’s audio, what was it trained for, and how does it perform on the calls we expect?”
That is a much better procurement question because it asks about the point where the customer experience can fail.
Test the language path with the calls you actually expect
A language claim deserves a language-level test. Build a small evaluation set from consented or safely created examples that reflect the job your agent will do. Include Twi phrases, English phrases, switches between them, names, locations, numbers, and the opening disclosure your callers need to understand.
Then inspect the whole path:
- Can speech recognition accurately capture what a caller says in Twi?
- Does synthesis pronounce the agent’s Twi response in a way your team considers usable?
- What happens when a caller switches languages halfway through a sentence?
- Can the transcript, audio, and call outcome be traced after the call?
- Does the system preserve consent, opt-out status, and an audit record?
A polished demo can conceal these questions because scripted phrases avoid the difficult parts. Real calls do not. A caller may repeat a delivery location, interrupt the agent, use a regional expression, or reject the call entirely. The evaluation should make room for those moments.
For a related example of why a Twi conversation needs its own technical model, see African Voice AI: Why Twi Needs Its Own Operating Model.
What Asenda Talk has built, and what remains gated
Asenda Talk is building a self-serve platform for voice agents with native Twi speech recognition and synthesis fine-tuned in-house. The aim is to give Ghanaian and African businesses a speech layer designed for the language their customers actually use, alongside agent configuration for persona, first message, and voice.
The platform also has a telephony lifecycle webhook pipeline for tracking call truth, consent and opt-out handling with an audit trail, metered per-minute billing, and admin secrets management designed to keep values write-only, masked, and environment-aware. Assistant runtime calling is Vapi-orchestrated.
It is in active early access. Feature parity with established voice-agent platforms is still in progress. Outbound calling is also gated behind an explicit telephony-provider decision that has not been made live. That boundary matters: a technical evaluation can assess the built speech and platform components today, but it should not assume a live outbound programme is ready until that provider decision is in place.
Ask for evidence before you turn on a call programme
By late afternoon, Kojo had stopped comparing logos and started writing an evaluation plan. He wanted sample audio, transcripts, language-switching cases, and a clear explanation of which system handled each step. He also wanted to know how a declined call, a consent record, and a completed call would appear in the audit trail.
That plan did not make a vendor choice for him. It gave his team a way to make one without confusing a capable orchestration layer for a language-ready speech system.
For voice programmes that contact customers, the check should extend beyond recognition quality. Confirm the consent path, opt-out handling, call outcome definitions, and the operator controls around real-money usage. A voice agent can sound natural and still create an operational problem if the call record cannot explain what happened later.
The next morning, Kojo’s recommendation was simple: run the Twi and English evaluation first, inspect the evidence call by call, and only then decide whether the platform belongs in the support desk’s next build.
Comments
No comments yet.