Model and Voice Settings
How to choose models and voices for Voice Agent in Standard and Realtime modes.
Overview
Users can choose personal model and voice preferences in Web App > avatar > Settings > Voice. A field left empty follows the assistant's configuration, else the platform default, and shows what it resolves to, for example Assistant default (Grok 2 TTS).
Every plan can use every active model; more capable models use more credits. See AI Models for the catalog, voice lists, and how a model is chosen.
The Language model (LLM) here is used for voice turns and is also your Default Chat model in the Web App header. A model picked in the header still overrides it for text chat. See Chat Model Selection.
Admins enable voice, recording policy, and Stream Response Text in the Admin Panel. See Enable Voice.
Standard Mode
Standard mode runs a chained pipeline: listen to speech → transcribe to text → generate the AI answer → read the answer aloud with TTS.
| Setting | Role | How to choose |
|---|---|---|
| Voice activity detection (VAD) model | Detects when the user starts and stops speaking | Keep Default unless you need to tune turn-ending behavior |
| Speech recognition (ASR) model | Transcribes speech to text | Choose a stronger model if transcripts are often wrong, especially for Vietnamese, accents, or noisy environments |
| Language model (LLM) | Generates the answer | Use a fast/cost-effective model for normal Q&A; choose a stronger model for reasoning, tool use, or long questions |
| Text-to-speech (TTS) model | Converts the answer into spoken audio | Choose based on voice quality, latency, language support, and credit cost |
| Voice | The spoken persona for Standard mode | Lists the voices of the TTS model in effect (yours or the default); changing the TTS model resets the voice |
| Speech speed | How fast TTS audio is spoken | Slow it down for careful listening; speed it up for quick operational use |
| Silence duration | How long to wait before ending the user's turn | Increase it if users pause often; decrease it for faster responses |
Realtime Mode
Realtime mode uses a native voice model to process live audio directly with very low turn latency. Because of that, Realtime has its own model and voice selection. Realtime does not use the Standard-mode TTS voice.
| Setting | Role | How to choose |
|---|---|---|
| Realtime model | Handles native live voice conversation | Choose a low-latency model for natural conversation; choose a stronger model when instruction following or tool use is more complex |
| Realtime voice | Native voice for the selected Realtime model | Lists the voices of the Realtime model in effect (yours or the default); changing the Realtime model resets the voice |
If the Realtime section does not appear, no Realtime model is active on the platform yet.
Credit Cost
Voice usage consumes credits based on the model and usage:
| Type | How it is charged |
|---|---|
| ASR | By audio duration |
| TTS | By character count or audio duration, depending on provider |
| Realtime Voice | By session duration |
A Realtime session starts only when your credits cover at least a minute on the selected model, and ends automatically when they run out.
See Credits & Usage to understand how credits are calculated.
If you are not sure what to choose, keep Default. Defaults are set by the assistant owner (or the platform) and are usually balanced for quality, speed, and credits.