Creating
In the 'AI Agent' section → 'AI Agent settings' tab → left menu 'All AI Agents':
Click the 'Create agent' button.
Specify the agent name.
Click the 'Save' button.
Configuration
Open agent settings by clicking on the block itself or the three-dot menu within it.
Fill in the fields or select from dropdown lists:
Tab General Settings:
Name — unique agent name for identification in the system.
Model — AI model that determines the agent's intelligence and capabilities:
- gpt-4o-realtime-preview — Flagship model for premium tasks with maximum accuracy and minimum latency.
- gpt-4o-mini-realtime-preview — Optimized version for balance of speed and cost, suitable for most tasks.
Voice — speech generation with unique characteristics:
- alloy — Neutral technological timbre,
- ash — Soft calm voice,
- ballad — Expressive melodic timbre,
- coral — Bright energetic voice,
- echo — Clear announcer timbre,
- sage — Calm confident voice,
- shimmer — Light airy timbre,
- verse — Rhetorical expressive voice.
Prompt — System instruction defining the AI agent's role, behavior, and communication style.
Tab Advanced settings:
The role of the first message — choose whose behalf the very first message in the dialogue with the user will be composed from.
- User — The message will imitate a reply from a person addressing the agent.
- System — The message contains instructions for the agent itself. This is the most important parameter. It defines the agent's role (for example, "You are a friendly support assistant for an online store"), its main tasks, rules of behavior, and communication style. This is a "meta-instruction" that the agent will remember throughout the dialogue.
- Agent — The message is the AI agent's first response. This is used to immediately start the dialogue with a specific phrase (for example, "Hello! How can I help you?").
The text of the first message — text that will be used depending on the selected Role of the first message.
Speed — response speed in the user interface.
Temperature — parameter controlling randomness and creativity of responses. From 0 to 1, recommended — 0.7.
Speech Detection Settings:
Type of speech detection — Algorithm for determining the beginning and end of speech (for example, based on sound energy or ML model).
- server_vad — mode where audio processing occurs on the server.
Delay before start (ms) — Waiting time after the first sound to ensure it's speech, not random noise.
Sensitivity threshold — Volume level above which sound is considered speech. The higher, the fewer false triggers from noise.
Duration of silence (ms) — Time of continuous silence after which the dialogue is considered completed.
Create a response automatically — If enabled, AI will start generating a response immediately after detecting the end of the user's speech, without additional commands.
Speech recognition:
Model — choice of algorithm for converting voice to text. Speed and accuracy depend on the model.
- gpt-4o-transcribe — model from OpenAI based on GPT-4o. Highest accuracy and context understanding. Best handles complex audio (accents, noise, specialized terms). The most advanced model.
- gpt-4o-mini-transcribe — lightweight and faster version based on GPT-4o Mini. Optimal for most standard tasks. Good balance of speed, cost, and quality.
- whisper-1 — classic open-source model from OpenAI. Universal operation, good quality for many languages.
Prompt — system prompt or context to improve recognition.
Language — language in which the user's speech is expected.
Noise reduction:
- Type of noise reduction— algorithm for filtering background noise.
near field — for quiet environments and closely positioned microphone (for example, call center).
far field — for noisy rooms and remote microphone (for example, smart speaker in a room). More aggressively suppresses noise.
- Type of noise reduction— algorithm for filtering background noise.
Tab Functions — add functions created in the 'AI agent' section on the 'Functions' tab. For more details, see 'AI Agent Functions'.
After filling in, click the 'Save' button.
Also see:
Was this article helpful?
That’s Great!
Thank you for your feedback
Sorry! We couldn't be helpful
Thank you for your feedback
Feedback sent
We appreciate your effort and will try to fix the article