---
title: "Fix silent, slow or interrupted audio"
description: "Separate microphone, playback, model and tool problems in a voice test."
url: "https://documentation.verbatik.com/troubleshooting/voice/"
---

# Fix silent, slow or interrupted audio

Separate microphone, playback, model and tool problems in a voice test.

## No microphone or no sound
1. Save the agent, choose an available voice model and start a fresh **Test → Audio** session.
2. Allow microphone access in the browser's site settings. Confirm that the intended microphone is selected at the operating-system level.
3. Check speaker output, volume and mute state. Click the call control yourself so the browser can allow audio playback.
4. Close another application that may hold the microphone. Use a headset to reduce speaker echo.
5. End the session before starting another one.

A missing microphone and denied microphone permission are different errors. If the microphone works but the selected provider cannot generate speech, preserve the displayed error and try an available voice from the selected model's catalog. A voice from another engine may not be compatible.

## Find where a slow turn waits
Compare a simple question with a question that uses a calendar or CRM. If only the latter is slow, review the tool outcome and integration response. The agent may use a short contextual filler while a tool runs; the filler does not confirm that the action succeeded.

Review [Analytics](/analytics/) and the session in [Call history](/call-history/). Keep the call ID, model, language, approximate time and whether a tool was involved. A missing latency measurement is not a zero-millisecond response.

For slow recognition or premature interruptions, test the same scenario in a quiet room before changing **Speech settings**. Change one threshold or timing setting at a time. Test both a short answer and a speaker who pauses mid-sentence. Extremely short pauses can make the agent interrupt a thoughtful caller.

## Read playback status correctly
**Playback interrupted** means the caller did not necessarily hear the complete response. **Generated · playback unconfirmed** is not a completed audio receipt. Check whether an action already succeeded before repeating the conversation or action.

## Frequently asked questions
**A text test passed. Is phone audio ready?** No. Text tests do not check microphone capture, speech quality, carrier audio or phone transfers.

**Can I compare every engine with automated suites?** Current suites support the Mistral pipeline. Use individual editor text/audio tests for the other managed realtime engines, and validate the engine on the intended deployment.

Continue with [browser testing](/web-testing/), [speech tuning](/speech-settings/) and [voice models](/voice-engines/).