A ten-minute Twi smoke test can reveal whether a voice platform recognizes and responds to a simple Twi request, or falls back to an English-first failure message. Run the same prompt, “yɛfrɛ me,” through every candidate platform before you spend time designing a call flow or moving a campaign into production.
At 4:40 p.m. in a small office in Accra, an illustrative developer named Kofi had four browser tabs open, a cold bottle of water beside his laptop, and one task left before a team review. The team wanted a voice agent that could handle Twi and English for callers. Kofi had already heard polished English demos from each vendor. Now he pressed record and said, “yɛfrɛ me.”
Platform one returned a response that matched the request. Platform two did the same. The next two answered, “Sorry, I didn’t get that.”
The risk was bigger than an awkward demo. If the team chose a platform based on English performance and a few Twi phrases in a scripted flow, callers could reach the live agent, speak naturally, and hit a dead end. The campaign might still show completed calls in a dashboard while the actual conversation had failed at its first meaningful turn.
Kofi did not need a long procurement exercise to see the difference. He needed a small test that exposed it.
Start with one ordinary phrase, spoken naturally
“yɛfrɛ me” is useful because it is short, ordinary, and easy to say at normal speed. It gives a platform little room to hide behind a carefully prepared English prompt or a demo script.
Record the phrase yourself, then try it through each platform’s speech input and agent runtime. Do not correct the transcript before sending it. Do not translate it into English for the model. The point is to observe what happens when a caller speaks Twi as they would on a real call.
Watch for three separate outcomes:
- Did the system recognize the spoken phrase in a form that preserves its meaning?
- Did the agent answer appropriately, rather than treating the phrase as noise?
- Did the voice sound capable of carrying a Twi reply that a caller can understand?
A single successful transcription does not settle the question. Speech systems can behave differently across voices, recording conditions, and phrasing. But a basic prompt that immediately becomes “sorry, I didn’t get that” gives you a clear signal: the platform needs more investigation before it sits in front of callers.
This is where “supports African languages” needs unpacking. A vendor may mean that a language can appear in an agent prompt. Another may mean the system can recognize and synthesize speech in that language. Those are different claims, with different consequences for a support desk or campaign team.
Add variation before you call it a language test
Kofi ran the same phrase again, this time more quickly. Then he added a short English continuation. He tried it from his phone, where the room noise and microphone differed from his laptop.
That small set of variations mattered. Real callers do not deliver a clean, isolated phrase once and wait for a score. They interrupt, switch languages, repeat themselves, and use the words that come naturally.
Keep the test narrow enough to finish in ten minutes:
- Say the core Twi prompt at a normal pace.
- Repeat it with a different speaker or device, if available.
- Add one short Twi-English switch.
- Ask the agent to respond, then follow up in Twi.
- Check both the transcript and the audio reply.
Write down the exact phrase, the result, and where it broke. “Poor Twi” is too vague to act on. “Recognized the phrase but replied in English” is useful. “The agent responded once, then failed after a mixed-language follow-up” is useful too.
For a deeper check on that second case, see What Happens When Twi-English Switching Sends Callers in Circles?. The language test should cover the conversation, not only the first transcript.
Separate language performance from call operations
A platform can pass the phrase test and still be a poor fit for a calling workflow. Language capability is one decision. Consent handling, opt-outs, call records, billing controls, and telephony readiness are separate decisions.
This distinction became clear when Kofi returned to his notes before the review. Two platforms had answered the Twi prompt. That narrowed the field, but it did not answer whether either could support the team’s real operating requirements.
Asenda Talk is built around native Twi speech recognition and synthesis fine-tuned in-house, alongside configurable voice agents and a Vapi-orchestrated assistant runtime. It also includes a telephony lifecycle webhook pipeline for call-truth tracking, consent and opt-out records, audit trails, masked environment-aware secret management, and metered per-minute billing with an operator-controlled real-money gate.
Those are capabilities to evaluate in the product today. Asenda Talk remains in active early access. Production outbound calling is gated behind an explicit telephony-provider decision that has not been made live. A Twi smoke test can validate speech behavior; it cannot turn an unavailable production telephony decision into a live calling program.
That caveat is worth putting in the evaluation document, alongside every result. A clear “not live yet” is more useful than a vague promise during vendor selection. For the operational side of respecting a caller’s stop request, read What Happens When a Caller Switches to Twi, Then Says Stop Calling Me?.
Make the smoke test part of every vendor review
The next morning, Kofi’s comparison was one page long. Each platform had the same prompt, the same test conditions, the response, the transcript, the reply language, and a note on follow-up behavior. The result was not a grand benchmark. It was enough to prevent an English demo from deciding a Twi conversation.
Use the same approach with the words your callers actually use. Keep the recordings. Retest after a model, voice, or prompt change. Then move to longer conversations only after the basic exchange holds.
The first caller should not be the person who discovers that the agent cannot understand “yɛfrɛ me.”
Comments
No comments yet.