> ## Documentation Index
> Fetch the complete documentation index at: https://dialnexa.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# How DialNexa Voice AI Works

> Learn how DialNexa Voice AI works from caller audio to Deepgram or Soniox transcription, OpenAI, Google, or Groq reasoning, tool calls, ElevenLabs or Cartesia speech, and call history records.

DialNexa Voice AI works by moving each call through a loop: receive audio, transcribe speech, decide the next response with the configured agent version, optionally call tools, synthesize speech, and store the result in Call History.

<Tip>
  If you can follow the loop, you can debug most calls without guessing. Guessing is dramatic, but logs are cheaper.
</Tip>

## The Runtime Loop

<Steps>
  <Step title="Caller audio arrives">
    Audio enters through a Plivo number, SIP trunk, web call, batch call, workflow call node, API call, or test call.
  </Step>

  <Step title="The transcriber listens">
    Deepgram or Soniox converts speech into text. The transcript is later shown in Call History as realtime and, where available, post-call text.
  </Step>

  <Step title="The agent decides">
    The published agent version supplies prompt, system prompt, language, LLM model, dynamic variables, functions, and safety settings.
  </Step>

  <Step title="Tools may run">
    Functions and integrations can end a call, book a calendar event, call an API, send a WhatsApp message, send an email, or trigger another configured action.
  </Step>

  <Step title="The voice speaks">
    ElevenLabs or Cartesia synthesizes the response. Audio Cache can reduce latency for repeated exact phrases.
  </Step>

  <Step title="Evidence is saved">
    Call History receives status, summary, transcript, recording URL, post-call analysis, transfer details, Audio Cache data, and metadata.
  </Step>
</Steps>

## Provider Work At Each Runtime Layer

| Layer            | Provider choices                                                                                                                                                         | Why users should care                                                                            |
| ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------ |
| Audio route      | Plivo, SIP trunking, web call.                                                                                                                                           | Changes number ownership, audio quality, routing, and whether phone network issues are involved. |
| Speech to text   | Deepgram or Soniox.                                                                                                                                                      | Changes transcript accuracy, language fit, turn timing, and Response Eagerness support.          |
| Reasoning        | OpenAI, Google, or Groq. Speech-to-Speech agents use a realtime model (OpenAI gpt-realtime-2 or Gemini 3.1 Flash Live) that replaces the separate transcriber and voice. | Changes instruction following, latency, structured behavior, and fallback strategy.              |
| Text to speech   | ElevenLabs or Cartesia.                                                                                                                                                  | Changes voice identity, pronunciation, streaming behavior, and cache compatibility.              |
| External actions | Wati, Resend, Custom Functions, webhooks.                                                                                                                                | Changes what the call or workflow can do outside DialNexa.                                       |

## Runtime Input Versus Saved Evidence

Do not confuse what starts a call with what proves it happened.

| Layer         | Examples                                                                                          | Where reviewed                                               |
| ------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------ |
| Input         | Recipient number, metadata, dynamic variables, workflow lead variables, selected outbound number. | Batch setup, workflow lead, API request, or test call modal. |
| Agent version | Prompt, model, voice, transcriber, welcome mode, functions, post-call fields, fallback settings.  | Agent builder and version history.                           |
| Call result   | Status, duration, sentiment, end reason, transcript, summary, recording, extracted fields.        | Call History and call detail page.                           |

## What Can Change A Reply

The same caller sentence can lead to different behavior when these settings change.

<CardGroup cols={2}>
  <Card title="Prompt and system prompt" icon="message-square" href="/agents/prompts-and-welcome-messages">
    Instruction wording decides goal, boundaries, escalation rules, and acceptable answers.
  </Card>

  <Card title="Dynamic variables" icon="braces" href="/agents/dynamic-variables">
    Caller-specific values can change greeting, eligibility, due date, location, or transfer destination.
  </Card>

  <Card title="Functions" icon="settings" href="/agent-settings/functions">
    Tools add actions the model can call during the conversation.
  </Card>

  <Card title="LLM and temperature" icon="brain" href="/voice-ai/llms-and-conversation-behavior">
    Model choice and temperature affect reasoning style and consistency.
  </Card>
</CardGroup>

## Debug By Layer

<AccordionGroup>
  <Accordion title="Audio sounds bad">
    Check recording quality, phone path, SIP trunk behavior, web microphone, and Denoising Mode.
  </Accordion>

  <Accordion title="Transcript is wrong">
    Check language, Deepgram or Soniox selection, background noise, and whether the caller spoke over the agent.
  </Accordion>

  <Accordion title="Transcript is right but answer is wrong">
    Check prompt, dynamic variables, functions, knowledge source, model family, and temperature.
  </Accordion>

  <Accordion title="Answer is right but late">
    Check Response Eagerness, Audio Cache, fallback LLM, function latency, and integration action placement.
  </Accordion>
</AccordionGroup>

## Related Reading

<CardGroup cols={2}>
  <Card title="Speech To Text" icon="file-text" href="/voice-ai/speech-to-text-and-transcription">
    Understand Deepgram and Soniox behavior.
  </Card>

  <Card title="LLM Behavior" icon="brain" href="/voice-ai/llms-and-conversation-behavior">
    Tune reasoning and fallback behavior.
  </Card>

  <Card title="Provider Selection Guide" icon="sliders-horizontal" href="/voice-ai/provider-selection-guide">
    Choose the complete provider stack.
  </Card>

  <Card title="Text To Speech" icon="volume-2" href="/voice-ai/text-to-speech-and-voices">
    Choose ElevenLabs or Cartesia.
  </Card>

  <Card title="Integrations" icon="plug" href="/dashboard-integrations/overview">
    Understand Wati, Resend, and workflow actions.
  </Card>
</CardGroup>