Build a voice agent that can follow the customer.
Language switching, clarification, and confirmed actions belong in the same conversation.
Humlet editorialA customer does not organize a request around a speech API. They explain the problem, switch to a more comfortable language, correct a detail, and expect the conversation to keep up.
A useful voice agent needs to follow those changes while keeping the difference between hearing a request and carrying it out clear.
Follow the language of the request.
A caller may greet the agent in English, describe a problem in another language, and use an English product name in the middle. Fixing the entire call to the greeting language can be the wrong assumption.
Design recognition and translation around this possibility. Test language changes within a turn as well as between turns, and make sure important terms survive the transition. A smooth greeting does not establish that the agent can understand the rest of the call.
Track details through corrections.
Imagine an order for three drinks that changes to two, with delivery at seven instead of six. The application needs the final quantity and time, plus enough context to know that the speaker corrected an earlier request.
A readable transcript helps, but the workflow should also track which values are confirmed. Repeat the relevant details before submitting an order. If the audio is unclear, ask about the uncertain item instead of making the customer start over.
Make the action a separate step.
Understanding “I’d like to reserve a table” is not the same as completing a reservation. A booking system can reject the requested time, a connection can fail, or the customer can change their mind.
The agent should state success only after the underlying action succeeds. Until then, it can acknowledge the request, check availability, and ask for confirmation. The spoken response and the application state need to tell the same story.
Give the conversation a graceful exit.
Some requests need a person. Preserve the relevant context for a handoff, explain what is happening, and avoid making the caller repeat information that was already captured correctly. Provide a way to recover when recognition or translation is unavailable.
Humlet’s interest in voice agents begins with multilingual understanding and a consistent speech layer. The application still owns its business rules, confirmation steps, and external actions. Evaluate the entire conversation, including failures, before describing it as an autonomous workflow.
Good voice automation connects understanding to confirmation, then to a verified action.
Building something that listens?
Explore the tools, or prepare a brief for your next speech integration.