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

# Configure Persona

> Fine-tune system prompts, STT speech engines, TTS vocal characteristics, and pronunciation dictionaries

The **Persona** tab allows you to configure your agent's character, vocal expression, and linguistic capabilities. Here you pair the brain (LLM) with ultra-low-latency hearing (STT) and expressive vocal speech engines (TTS).

![Persona Configuration Tab in Vaani Dashboard](https://mintcdn.com/vaaniai/1z97x591P_yo5Lyb/images/screenshots/persona-agent.png?fit=max\&auto=format\&n=1z97x591P_yo5Lyb\&q=85\&s=6f911529a4827d8d85a848e2268ee5f5)

## Video Walkthrough

[https://www.loom.com/embed/830ca7036f68471c94fb1a819d700285](https://www.loom.com/embed/830ca7036f68471c94fb1a819d700285)

***

## 1. Ears (Speech-to-Text / STT)

The STT engine determines how accurately and quickly your agent hears caller speech over cellular or VoIP telephone connections.

**Primary Provider**

We recommend **Deepgram Nova-2** for real-time conversational telephony:

* **Latency**: Sub-200ms streaming transcription.
* **Acoustic Robustness**: Handles background traffic, static, and speakerphone echoes gracefully.
* **Multilingual Support**: Supports Indian English, Hindi, and regional dialects.

**Fallback Provider**

Configure **AssemblyAI** or **Whisper** as a secondary fallback:

* If the primary provider experiences network degradation or downtime, Vaani switches automatically with zero dropped calls.

**Keyword Boosting**

Improve recognition accuracy for hard-to-transcribe terms:

* Company names (e.g. *Aangan Studio*, *VaaniVoice*)
* Technical acronyms (e.g. *KYC*, *SIP*, *RTGS*, *VoIP*)
* Regional locations (e.g. *Kothrud*, *Hinjewadi*, *Bandra*)

***

## 2. Brain (Large Language Model)

Configure the intelligence engine driving conversational decisions:

| Parameter | Recommended Setting | Description |
| - | - | - |
| **Model** | `GPT-4o`, `Claude 3.5 Sonnet`, `Llama 3.3 70B` | Select the model architecture best suited to your task complexity |
| **Temperature** | `0.2` – `0.5` | Lower values ensure factual, consistent answers; higher values produce creative, relaxed dialog |
| **Max Tokens** | `80` – `150` | Limits turn length to prevent the agent from delivering unnaturally long monologues |

***

## 3. Mouth (Text-to-Speech / TTS)

Select the vocal personality that reflects your brand:

* **Sarvam Bulbul:v3**: Exceptional naturalness for Indian accents, Hindi, Marathi, Tamil, and Hinglish.
* **ElevenLabs Multilingual v2**: High-definition expressive emotional voices.
* **Deepgram Aura**: Ultra-low time-to-first-byte (TTFB) for lightning-fast turn-taking.

### Pronunciation Dictionaries

Synthetic voices occasionally mispronounce specialized words, brand names, or abbreviations. You can provide explicit phonetic respellings:

```json theme={null}
{
  "Vaani": "VAA-nee",
  "Aangan": "AAN-gun",
  "Pune": "POO-nay",
  "cURL": "see-url"
}
```

***

## 4. Voice Lab Integration

Want to test custom voice clones or audition different voice actors before assigning them to your agent? Head over to the [Voice Lab](https://docs.vaanivoice.ai/getting-started/voice-lab.md) to clone your own voice or audition pre-licensed voice actors.

***

## Next Steps

> **[Training & Knowledge Base](https://docs.vaanivoice.ai/getting-started/agent-config/training.md)**
>
> Ground your persona with company documents, FAQs, and API tools

> **[Experience Tuning](https://docs.vaanivoice.ai/getting-started/agent-config/experience.md)**
>
> Configure ambient sound, interruption sensitivity, and speech delays


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.