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

# AI Configuration (Deep Dive)

> Configure Burki voice AI assistants with LLM, TTS, STT, fallback providers, voice settings, speech detection, and prompt behavior.

<Callout type="info">
  This is where you give your assistant its brain, voice, and ears! Configure how your AI thinks, speaks, and listens.
</Callout>

***

## The Tabbed Interface

When creating or editing an assistant, you'll see three main tabs:

<Tabs>
  <Tab label="LLM (Language Model)">
    ### What is the LLM?

    The LLM (Large Language Model) is the "brain" of your assistant. It understands what callers say and generates smart, helpful responses.

    ### Key Settings

    <Checklist>
      * **Provider:** Choose from OpenAI, Azure OpenAI, Anthropic, Gemini, xAI, Groq, custom HTTP, or custom WebSocket.
      * **API Key:** Paste your provider's API key (get it from their dashboard).
      * **Model:** Pick the model or deployment name (e.g., GPT-4o, Claude 3.5 Sonnet, Gemini 2.5 Flash, Groq Llama, or an Azure deployment).
      * **Base URL:** Set a custom API endpoint for Azure OpenAI and custom/OpenAI-compatible providers.
      * **Temperature:** Controls creativity. Lower = more predictable, higher = more creative. <br /> <Callout type="tip">0.7 is a good starting point!</Callout>
      * **Max Tokens:** Limits response length. <br /> <Callout type="tip">1000 tokens ≈ 250 words. Good for phone calls.</Callout>
      * **System Prompt:** Set your AI's personality and instructions (e.g., "You are a friendly assistant...").
      * **Greeting Message:** The first thing callers hear.
      * **Fallback Providers:** Add up to 3 backup LLMs for reliability.
    </Checklist>

    <Accordion title="How do Fallbacks Work?">
      If your main LLM provider fails (e.g., rate limit, downtime), the system automatically tries your backup providers in order. This keeps your assistant reliable—even if one provider has issues.
    </Accordion>
  </Tab>

  <Tab label="TTS (Text-to-Speech)">
    ### What is TTS?

    TTS turns the AI's text responses into a natural-sounding voice. Choose from multiple providers to match your quality, speed, and language needs.

    ### Quick Provider Guide

    <CardGroup cols={2}>
      <Card title="⚡ ElevenLabs" href="/tts-providers/elevenlabs">
        **Premium Quality & Multilingual**

        70+ languages, voice cloning, advanced controls
      </Card>

      <Card title="🚀 Deepgram Aura" href="/tts-providers/deepgram">
        **Ultra-Low Latency**

        vendor-reported low latency, phone-optimized
      </Card>

      <Card title="🎭 Inworld.ai" href="/tts-providers/inworld">
        **AI-Powered Emotions**

        Multilingual with emotional markup
      </Card>

      <Card title="🎯 Resemble AI" href="/tts-providers/resemble">
        **Custom Brand Voices**

        Unlimited voice creation
      </Card>

      <Card title="🧠 OpenAI TTS" href="/tts-providers/openai">
        **OpenAI-Native Voices**

        `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts`
      </Card>

      <Card title="☁️ Azure / Cartesia" href="/tts-providers">
        **Enterprise and Multilingual**

        Azure Speech, Cartesia, plus Kokoro, Murf, Soniox, and Uplift
      </Card>
    </CardGroup>

    ### Configuration Overview

    <Checklist>
      * **Provider:** Choose based on your needs ([Full Comparison](/tts-providers))
      * **Model:** Select quality/speed balance (varies by provider)
      * **Voice:** Pick from available voices or use custom voices
      * **Language:** Select for multilingual providers
      * **Voice Controls:** Fine-tune stability, similarity, style ([Voice Tuning Guide](/tts-providers/voice-tuning))
      * **Advanced Settings:** Encoding, sample rate, streaming options
    </Checklist>

    <Accordion title="Quick Setup by Use Case">
      **Phone Calls:** ElevenLabs Flash v2.5 or Deepgram Aura-2 for speed

      **Customer Service:** ElevenLabs Bella or Inworld Ashley for warmth

      **Multilingual:** ElevenLabs v3, Cartesia Sonic 3, or Inworld `inworld-tts-1.5-mini` with native voices

      **Brand Voice:** Resemble AI with custom voice training

      → [See detailed recommendations](/tts-providers#quick-start-guide)
    </Accordion>

    <Card title="📚 Complete TTS Documentation" icon="book">
      **Dive Deeper:**

      * [TTS Provider Comparison](/tts-providers) - Choose the right provider
      * [Voice Tuning Guide](/tts-providers/voice-tuning) - Optimize voice settings
      * [Best Practices](/tts-providers/best-practices) - Performance & cost optimization
      * [Troubleshooting](/tts-providers/troubleshooting) - Fix common issues
    </Card>
  </Tab>

  <Tab label="STT (Speech-to-Text)">
    ### What is STT?

    STT converts what the caller says into text for the AI to understand.

    ### Key Settings

    <Checklist>
      * **Provider:** Choose Deepgram, ElevenLabs, Azure, Uplift, Speechmatics, Telnyx, or Soniox.
      * **Model:** Choose a provider model such as `nova-3`, `flux-general-en`, `scribe_v2_realtime`, `standard`, `deepgram/nova-3`, or `stt-rt-v4`.
      * **Language:** Choose the spoken language (or enter a custom code)
      * **Advanced Timing Controls:** Fine-tune speech detection (see [STT Providers](/stt-providers) for details)
      * **Punctuation/Interim Results/Smart Format:** Make transcripts easier to read and use
      * **Audio Denoising:** Remove background noise for better accuracy (uses RNNoise)
      * **Keywords/Keyterms:** Boost recognition of important words (e.g., company names)
    </Checklist>

    <Accordion title="Troubleshooting STT">
      * If the AI misses words, try enabling denoising or adding keywords
      * For non-English calls, set the correct language code
      * Adjust silence settings if the AI cuts callers off too soon or waits too long
    </Accordion>
  </Tab>
</Tabs>

***

## Best Practices

* **Start simple:** Use recommended defaults, then experiment with advanced settings
* **Test with real calls:** Try different voices, models, and fallback setups
* **Document your changes:** Keep track of what works best for your use case

***

## What's Next?

<CardGroup cols={2}>
  <Card title="📞 Call Management" href="/call-management" icon="phone">
    Configure interruption handling, timeouts, and conversation flow control
  </Card>

  <Card title="🎙️ STT Advanced Settings" href="/stt-providers" icon="microphone">
    Fine-tune speech detection timing and audio processing
  </Card>

  <Card title="🔊 TTS Provider Details" href="/tts-providers" icon="volume-high">
    Deep dive into voice options and audio optimization
  </Card>

  <Card title="🛠️ Tools & Custom Actions" href="/tools" icon="wrench">
    Built-in call actions plus custom integrations with APIs, Python functions, and AWS Lambda
  </Card>
</CardGroup>

<Callout type="tip">
  If you get stuck, check the tooltips in the UI or visit the [FAQ](/faq).
</Callout>