Skip to main content
After creating your initial voice agent, you can customize its persona to match your brand’s voice and operational requirements. Fine-tuning your agent’s persona controls how it greets callers, how accurately it hears speech, its vocal tone and emotional expression, its memory across calls, and the external tools it can execute during live conversations.

​
Video Walkthrough

​
Persona Configuration Pillars

Customizing an agent consists of four main areas:
  1. Identity & Greeting — Defining who the agent is, its initial opening line, speech delay, and conversational pacing.
  2. Senses & Capabilities (AI Pipeline) — Choosing the best Speech-to-Text (STT), Large Language Model (Brain), and Text-to-Speech (TTS) voice engines.
  3. Tools & Integrations — Enabling the agent to look up customer data, book appointments, or trigger APIs via cURL integrations mid-call.
  4. Context & Memory — Giving the agent recall of past conversations with the same caller or shared knowledge across agents.

​
Step-by-Step Customization

1

Configure Identity & Greeting

Open Agent Config → select your agent → navigate to Persona → Identity:
  • System Prompt: Refine the instructions, guidelines, and behavioral traits for your agent. Use curly braces like {customer_name} or {company_name} for dynamic metadata variables.
  • Greeting Message: Specify the exact first words the agent says when the call connects.
  • Speech Delays & Interruption: Configure agent_speech_delay (how soon it speaks), silence tolerance before responding, and whether the caller is allowed to interrupt the greeting (interruptible: true).
  • Let User Speak First: Enable this if you prefer the agent to wait for the caller to say “Hello?” before delivering its opening line.
2

Set Up Ears (Speech-to-Text / STT)

Go to Senses & Capabilities → Ears:
  • Primary Provider: We recommend Deepgram for ultra-low latency and superior conversational transcription.
  • Fallback Provider: Configure AssemblyAI (or another supported STT engine) as a fallback in case the primary engine encounters network anomalies.
  • Keyword Boosting & Special Words: Add difficult or phonetically ambiguous terms (such as brand names, technical jargon, or place names like “Mississippi”) so the STT engine transcribes them accurately.
3

Select Brain (LLM) & Mouth (Voice / TTS)

Fine-tune the intelligence and voice of your agent:
  • LLM Engine: Choose the underlying language model and temperature based on your need for creative dialog vs. strict procedural accuracy.
  • Voice & TTS: Select a voice model suited to your brand’s regional accent, gender, and language.
  • Emotion & Pace: Adjust speaking rate and emotional tone (e.g., cheerful, empathetic, or authoritative).
  • Pronunciation Dictionaries: Provide custom phonetic respellings for words or acronyms the text-to-speech engine might otherwise mispronounce.
4

Add Custom Tools via cURL

Equip your agent to perform actions while on the call:
  • Paste a standard cURL request for your webhook or REST endpoint (e.g., check account balance, verify OTP, or schedule a calendar invite).
  • Provide clear natural-language descriptions of when the agent should trigger the tool.
5

Enable Agent Memory

Activate conversational recall under the Memory settings:
  • Previous Calls Memory: Enables the agent to remember context, preferences, and action items from prior calls with the same phone number.
  • Cross-Agent Memory: Allows context to persist even if a customer speaks with different agents across multiple departments.

​
Troubleshooting

Adjust the silence tolerance and user speech delay settings in your greeting and persona configurations. Increasing the pause threshold by 200–500ms prevents the agent from cutting in while the caller takes a breath.
Add the problematic terms to the Keyword Boosting list in your STT configuration, or add them to the Pronunciation Dictionary with their phonetic equivalents.
Yes. Edits saved in the dashboard apply to all new calls immediately. In-flight calls will finish using the configuration with which they were initiated.
Verify that the tool’s natural-language description clearly explains what triggers it and what arguments it requires. Also ensure that the external endpoint returns a response within 2–3 seconds so the conversation doesn’t stall.

​
Next Step

Voice Lab

Browse, audition, and swap realistic AI voices across providers like Cartesia and ElevenLabs