Skip to content

Voice debug monitor

Live debugging

Use the Voice Debug Monitor when you need evidence from a real call: what the caller said, what the agent heard, what context was sent, and when tools ran.

Debug the call path before rewriting prompts or changing multiple settings.

Use the Voice Debug Monitor when you need to see what happened during a live or test call while it is still unfolding. This is the fastest way to understand why the callback assistant asked the wrong question, missed a caller detail, or triggered a tool too early.

Where to find it

You can reach it in two places:

  • from the agent editor when the Debug step is available
  • from the agent actions menu on an existing voice agent

The page title in the product is Voice Debug Monitor.

What the monitor is for

Use it to inspect:

  • what the speech-to-text system actually heard
  • what the assistant said back
  • whether the backend debug stream is active
  • when tools executed
  • whether Pre Processing caller lookup saved context
  • whether knowledge base content, conversation facts, or conversation history were injected
  • which language-related prompt variables and STT language candidates were active

For the callback assistant, this is the best place to verify that the agent collects the name, company, reason, and callback number in the right order.

Before you start

The monitor is most useful when:

  • the agent is already assigned to a phone number
  • you are placing a real or controlled test call
  • Debug enabled is turned on for the voice agent

If the page loads but nothing useful appears, check whether the Debug enabled switch is on and whether the backend state shows Backend Debug On.

What you will see

The debug page is organized around live call activity.

Status badges

At the top of the page you will see:

  • connection status
  • Backend Debug On or Backend Debug Off
  • the Debug enabled switch
  • the Clear timeline button

The Why updates can pause hint is expected. The stream may update slightly behind real time, but finalized speech, tool executions, and knowledge hits should still be reliable.

Call Setup

The Call Setup section shows what happened when the call started:

  • the greeting text,
  • whether the greeting was non-interruptible,
  • any greeting timeout,
  • the executing agent,
  • Pre Processing caller lookup status and saved prompt context, when configured.

Use this section when you need to confirm that the greeting played correctly or that {pre_call_context} was available before the first agent reply.

Pre Processing events can show lookup states such as Running, Saved, Failed, or Skipped. Anonymous caller numbers may be skipped.

Caller Speech

This panel shows:

  • the current partial transcript
  • finalized utterances
  • silence gaps between utterances

Use this first when the callback assistant captures the wrong caller name, company, or phone number. If the speech transcript is wrong here, the problem is usually in recognition language, phrase list, or silence timing, not in the prompt.

Assistant Response

This panel reconstructs the assistant response from streaming chunks.

Use it to check:

  • whether the assistant stayed on script
  • whether it asked only one question at a time
  • whether it repeated itself
  • whether it responded too early or too late

For the callback assistant, a healthy sequence looks like:

  1. greeting
  2. caller name
  3. company name
  4. reason for the call
  5. callback number confirmation
  6. tool execution
  7. clean closing line

Prompt context and language variables

The monitor can show prompt-context details such as persona, default language, detected spoken language locale, active STT locales, allowed STT locales, and conversation fact summary.

Use this when:

  • a multilingual agent responds in the wrong language,
  • speech recognition appears to use the wrong language candidates,
  • a fact was collected but does not seem to appear in a later prompt,
  • you need to confirm what the assistant knew at a specific point in the call.

Debug snapshots are point-in-time. If a fact is stored after a prompt was built, it may appear on the next assistant turn rather than the previous one.

Conversation facts and history

The monitor can show conversation fact summary events and conversation history events. Use these to verify what context the agent had at a specific point in the call.

This is especially useful when:

  • the agent forgets a detail that was already stored,
  • a post-call summary does not include expected facts,
  • you are testing whether {conversation_history} and fact memory are actually available.

Knowledge Base query details

Knowledge Base events can include query transformation details, confidence, reasoning, and snippets. Use this when the agent answers from documents but the retrieved context looks surprising.

If the transformed query looks wrong, fix the knowledge-base or query-transform instructions before changing unrelated voice settings.

What to check for the callback assistant

During a test call, watch for these specific signals:

The assistant skipped a question

If the assistant jumps from greeting directly to ending the call or triggering a tool, check:

  • the system prompt conversation flow
  • the Email Tool or API tool execution description
  • whether the conversation history is being included

The assistant forgot something the caller already said

Check both:

  • Conversation history limit is set in Technical
  • {conversation_history} is still present in the prompt

If either is missing, the assistant will sound stateless even when the rest of the flow is correct.

The assistant triggered a tool too early

This is one of the highest-value debug checks. If the Email Tool or API tool runs before all callback details are complete, the execution description is too loose or the prompt does not clearly gate the action.

For the callback assistant, the tool should not run until:

  • caller name is known
  • company name is known
  • reason for the call is known
  • callback number is confirmed

The company name or caller name is wrong

If the monitor shows the wrong transcription, improve:

  • recognition language selection
  • active or allowed STT locales
  • phrase list entries
  • silence timing if the model is clipping words

Caller lookup did not provide context

Check Call Setup and the Pre Processing card. If the lookup failed, was skipped, or stored an unexpected value, fix the Pre Processing configuration before rewriting the agent prompt.

The assistant sounds repetitive

If the responses repeat or loop:

  • tighten the conversation flow in the prompt
  • keep one question per turn
  • reduce temperature if answers are drifting

Use this order every time:

  1. place a test call
  2. watch Caller Speech
  3. watch Assistant Response
  4. confirm tool execution timing
  5. make one change only
  6. place the same test call again

Do not change speech settings, prompt logic, and tool execution rules all at once. You will not know which change fixed the issue.

Common callback-assistant checks

Name captured incorrectly

Start with Caller Speech and phrase list updates.

Callback number not confirmed

Check the prompt flow. The assistant should explicitly ask whether {caller_phone_number} is the correct callback number.

Email never sends

Check the tool execution description and the debug stream for the moment when the tool should have fired.

Email sends with missing fields

The tool fired too early. Tighten the trigger logic before changing the email template.

Read this next

  • Common Issues if the assistant behavior is still unstable
  • Agent Options for the prompt and technical settings behind the debug stream
  • Call History to review the finished result after the call ends