The '70+ languages' feature list tells you nothing about how a Twi phone call actually sounds: here's a demo checklist: Twi word error rate, code-switching, and whether the first message sounds like someone from Accra.

A Twi demo should be judged on recognition accuracy, natural speech, code-switching, and how the call behaves when a person changes their mind. A long language count does not show whether a caller in Accra will be understood, or whether the agent’s opening line will sound credible enough to keep them on the line.

Bring a test script before you book the demo

Do not accept a vendor’s default prompt as the test. Bring a short script based on calls your team already handles: a payment reminder, appointment confirmation, delivery update, or support triage.

Include names, place names, amounts, dates, and phrases your callers use. A test such as “Medaase, me din de Ama. Mepɛ sɛ meka me appointment no ho asɛm” is more useful than a clean, isolated sentence read slowly into a microphone.

Prepare at least 20 utterances across three groups:

• Twi spoken at a normal conversational pace.
• English terms commonly used inside Twi sentences, such as mobile money, reference number, account, delivery, or WhatsApp.
• Messy real call inputs: interruptions, corrections, background noise, and people repeating themselves.

Ask for the transcript after every test. If the platform cannot show you what it heard, you cannot assess where the conversation went wrong.

Measure Twi word error rate on your own phrases

Word error rate, or WER, compares what a speaker said with the transcript produced by speech recognition. Lower is better, but a vendor’s aggregate figure only matters if the test data resembles your calls.

Ask how the number was measured. Was it Twi spoken by Ghanaian callers, or a small internal test set? Were speakers using phone audio? Did the evaluation include common English insertions? A result measured on clean recordings can look very different on a mobile call with traffic, fan noise, or a weak network.

You do not need a speech research team to run a useful check. Record or read your 20 test utterances, save the expected transcript, then compare it with the system output. Mark three kinds of mistakes:

• Meaning-changing errors, such as a wrong amount, date, name, or response to a yes or no.
• Routing errors, where the agent misses the reason for the call.
• Minor spelling or segmentation differences that do not change meaning.

For a payment workflow, a recognition error around “Friday,” an amount, or a reference number matters more than imperfect punctuation. Keep a separate tally for those high-risk fields.

Asenda Talk uses Twi speech recognition and synthesis fine-tuned in-house. That gives teams a reason to test Twi performance directly rather than treating a language label as proof. The platform remains in early access, so test the exact use case you plan to run instead of assuming feature parity with established voice-agent platforms.

Force a Twi-English switch in the middle of the call

Code-switching is where polished demos often fail. A caller may begin in Twi, give an English account number, ask for a callback “around 3pm,” then return to Twi to explain the problem. Your agent needs to retain the intent and respond appropriately, rather than restarting the flow or answering a different question.

Build switches into the script on purpose. Try phrases such as:

• “Mepɛ sɛ me reschedule the delivery for Friday.”
• “Medaase, but my reference number no yɛ 4729.”
• “Can you call me later? Me wɔ meeting mu.”

Listen for more than transcription. Does the agent understand the combined meaning? Does it keep the correct language in its response? If it asks the caller to repeat a clear phrase, test whether the second turn restores context or creates a loop.

This is especially important for customer service and collections calls, where a caller can shift from a routine response into a follow-up question. For a closer look at that failure mode, read What Happens When Twi-English Switching Sends Callers in Circles?.

Judge the first message as a phone moment

The first message carries more weight than a polished voice sample on a website. On a phone line, people decide quickly whether the speaker sounds familiar, relevant, and safe to continue with.

Provide the first line you would actually use. Include your business name, the reason for the call, and a simple route to decline. Then ask several Ghanaian Twi speakers to listen without telling them which vendor produced it.

Listen for pronunciation, pacing, pauses, and intonation. Does the voice flatten a name or place name? Does it pause in the wrong place? Does the greeting sound like it was assembled word by word? “Someone from Accra” is not a technical metric, but it is a practical usability test: would a local caller stay long enough to hear the purpose of the call?

Test both a Twi opening and a bilingual opening if your customers expect one. A voice can sound natural in a single sentence and become awkward when it moves between languages.

Test consent, stop requests, and the call record

A convincing voice means little if the call cannot be governed safely. During the demo, interrupt the agent with a clear stop request in Twi and English. Confirm that it acknowledges the request and that the record captures it.

Ask to see the audit trail: consent status, opt-out status, timestamps, and the final call outcome. Then ask what happens to a scheduled retry after an opt-out. The answer should be visible in the platform’s call controls, not dependent on an operator remembering a manual step. What Happens When a Caller Switches to Twi, Then Says Stop Calling Me? covers the operational risk in more detail.

Asenda Talk includes consent, opt-out, and audit tracking for every call, plus a telephony webhook pipeline intended to track call truth across the lifecycle. Outbound calling is currently gated behind an explicit telephony-provider decision that has not been made live. Treat any outbound demonstration as an evaluation of the available platform components, not a promise that a production outbound campaign can launch immediately.

Leave with a scored decision

Score each vendor on your own test set: transcript accuracy, meaning-changing errors, code-switch recovery, first-message naturalness, stop-request handling, and record visibility. Write down the conditions of the test, including phone versus browser audio and the exact prompt used.

Your next action is simple: send the same 20-utterance script to every platform you are considering, ask for transcripts and call records, and have two Twi-speaking reviewers score the opening message blind. That comparison will tell you more than a feature list ever can.