Make a voice assistant easier to understand and interrupt

A good voice experience helps the caller understand the answer and finish a task. It should not require the caller to mistake the assistant for a person. Evaluate intelligibility, responsiveness and truthful behavior together.

Begin with the words

Short spoken sentences are easier to follow than long paragraphs. Ask one question at a time when collecting details. Read back important names, numbers or selections in a form the caller can correct. Avoid repeated introductions and long transfer announcements that delay the next useful step.

Use approved business facts. A pleasant delivery does not make an invented price or unavailable appointment acceptable.

Evaluate the selected voice in context

Listen to a realistic exchange rather than only a voice preview. Include the business name, local place names, dates and numbers. Test each language the deployment actually needs with people able to assess it. No voice catalog establishes native-quality output for every language or speaker.

Compare the actual provider costs and billing units. Do not assume higher quality always has negligible cost or that more voice choices mean better conversations.

Test pauses and interruptions

The system must distinguish a caller finishing a thought from pausing mid-sentence. It also needs to handle a caller speaking during an answer. LiveKit's turn-handling guide explains the distinction between turn detection and interruption behavior.

For Burki, test the selected runtime and configuration. Try a correction halfway through an answer, a short acknowledgment, background speech and a longer pause. Observe when playback stops and whether the next answer uses the correction. A detector name or prompt instruction is not proof of caller-facing quality.

Review with consistent evidence

Measure response onset from the same reference point across tests. Distinguish generated audio from audio heard by the caller. If recording is enabled, use the appropriate permissions, disclosure and retention configuration; otherwise retain written observations and available session evidence.

Ask reviewers whether the assistant was understandable, whether interruptions worked and whether the result was correct. Do not ask them merely to guess whether the speaker was human.

The voice evaluation checklist combines these audio observations with action reliability and billing review.