Speech recognition and synthesis built specifically for Twi can preserve the language a customer already uses, instead of pushing them into English to complete a support call. That matters because a voice agent’s useful test is not whether it can produce a reply, but whether it can follow a real customer through a natural Twi and English conversation.
The moment a customer changes languages for the machine
A support lead hears the pattern during call review. The customer begins in Twi, explains the issue comfortably, then pauses when the bot responds and switches to English. The agent continues. The call may even reach an answer. Yet the language change is evidence: the customer is adapting to the system.
That adaptation can hide inside a successful transcript. A generic multilingual API may recognize familiar English terms, then lose meaning when a caller returns to Twi, mixes in a product name, or uses the phrasing they would use with a person at the desk. Synthesis creates a second test. If the reply sounds unnatural or breaks the rhythm of the exchange, the customer has more work to do before they can trust it.
For a Ghanaian support desk, “the call completed” is too weak a standard. Review where customers change language, repeat themselves, abandon a question, or receive a response that misses the point. Those moments show where the voice layer is setting the terms of the conversation.
A familiar technical failure: the wrong unit at the boundary
On September 23, 1999, NASA lost the Mars Climate Orbiter during its attempt to enter orbit around Mars. The spacecraft had been developed with Lockheed Martin Astronautics in Denver, while navigation work was handled at NASA’s Jet Propulsion Laboratory in Pasadena, California. NASA’s investigation found that one system provided impulse data in pound-seconds while another expected newton-seconds.
The spacecraft did not fail because engineers lacked a navigation system. It failed because two parts of the system used different assumptions at the boundary between them. By the time the mismatch mattered most, the mission could not recover.
NASA documented the incident in the Mars Climate Orbiter Mishap Investigation Board Phase I Report. The lesson travels well beyond spaceflight: a component can appear capable on its own and still fail the real task when its assumptions do not match the environment around it.
Language systems have boundaries too. A generic multilingual model may accept Twi audio and return a response. The support lead still needs to ask what happens at the points that carry meaning: local pronunciation, code-switching, names, common customer phrasing, and the cadence of a spoken answer. If the system treats Twi as an add-on rather than a language it has been trained and evaluated to handle, those boundaries deserve close testing.
What changes when Twi is the evaluation target
Asenda Talk is being built around native Twi speech recognition and synthesis fine-tuned in-house. The point is to evaluate Twi as the working language of the call, including the places where Twi and English naturally meet, rather than treating English performance as a proxy.
A useful test starts with recordings and scenarios your support team already recognizes. Include callers who stay in Twi, callers who move between Twi and English, names that matter to your operation, and the questions that produce the most follow-up work. Listen for more than word accuracy. Check whether the agent identifies intent, keeps the right context, responds in an appropriate voice, and gives the caller a clear next step.
Asenda Talk lets early-access teams configure an agent’s persona, first message, and voice, then inspect the call lifecycle through webhook events and call-truth tracking. That makes it possible to compare what the customer heard with what the system recorded. Consent, opt-out handling, and an audit trail should be part of the same review, especially before a team considers contact at scale.
The platform uses Vapi for assistant runtime today. More languages are in progress. It is still in active early access, and outbound calling remains gated behind an explicit telephony-provider decision that has not been made live. Those limits are important when planning a pilot.
Run a language-fit pilot before expanding call volume
Start with a narrow use case: appointment reminders, delivery updates, intake questions, or a contained support queue. Define a small set of calls where a Twi-speaking customer should never need to switch to English merely to be understood.
Then review the exceptions. Was the recognition wrong? Did the reply use the wrong language? Did the agent lose context after a switch? Did the call record show a different outcome from what the customer experienced? The same discipline applies to campaign reporting, where a completed technical event does not automatically mean a completed customer interaction. What Does “Completed” Actually Mean in a Voice Campaign Dashboard? explores that distinction.
Mars Climate Orbiter shows why an interface assumption deserves attention before it becomes operational fact. For a Twi voice agent, the practical equivalent is simple: test the language your customers speak when they are most comfortable, then use those calls to decide what the system is ready to handle.
Comments
No comments yet.