Voice configuration
By the end of this page you will have your agent’s voice fully configured — its voice and language, how it listens and speaks, and how it handles calls.
Prerequisites
- Starter Voice requires a Starter or higher plan.
- An AI Agent with Voice enabled (Modes).
- Admin or Developer role.
The Voice Configuration panel appears on the Modes tab when Voice is enabled.
Configuring voice makes the agent able to hold a spoken conversation. To let real customers phone it, give the agent a phone number — see Telephony.
Voice and language
Choose a TTS model
Pick the text-to-speech model: Google Chirp 3: HD, Google Neural2, OpenAI TTS-1, or OpenAI TTS-1 HD.
Choose the language
Select the language. The list adapts to the model — Chirp 3: HD supports 50+ languages; Neural2 and OpenAI support smaller sets.
Choose the voice
Pick the voice, which is also model-aware. Chirp 3: HD offers 30 named voices (with gender); OpenAI offers Alloy, Nova, Shimmer, Echo, Fable, and Onyx; for Neural2 the voice is selected automatically from the language and gender preference.
Set the speech speed
Adjust Speech Speed from 0.25× (very slow) to 2.0× (fast); 1.0× is normal.
Pipeline mode
The pipeline determines how the agent turns speech into a response and back.
| Pipeline | How it works |
|---|---|
| STT → ReAct → TTS | Separate transcription, reasoning, and speech synthesis — best accuracy and control. |
| Native Realtime (Beta) | An end-to-end speech model (for example, GPT-4o Realtime) — lowest latency, most natural conversation. |
Call behavior
Barge-in and DTMF
Enable Barge-In to let the caller interrupt while the agent is speaking. Enable DTMF Keypad to accept phone-keypad input (for example, “press 1 for…”).
Timeouts and duration
Set Silence Timeout (ms) (500–10,000) — how long to wait for the caller before prompting — and Max Call Duration (sec) (60–3,600).
Welcome message
Set the Welcome Message the agent opens with, for example “Hello! How can I help you today?”.
Call recording
Enable Call Recording only where appropriate — it requires explicit user consent (HIPAA compliance).
Messages
Expand Messages to customize what the agent says in common situations: the Fallback, Goodbye, Hold, and Transfer messages.
Transfer settings
Expand Transfer Settings to configure hand-off to a human:
- Human Transfer Number — where to route the call.
- Transfer on Failure — hand off after repeated failures.
- Max Consecutive Failures (1–10) — how many failures trigger a transfer.
Voice persona
Expand Voice Persona to shape how the agent sounds and what it says while working:
- Response Style — Conversational, Concise, or Detailed.
- Voice Gender — Female, Male, or Neutral.
- Max Response Sentences (1–10) — keeps spoken replies short.
- Pitch (−20 to +20) and Volume Gain dB (−10 to +10).
- Filler phrases — what the agent says while thinking or running tools: Thinking, Tool Call, Long Tool Call, Error, No Result, and a Periodic Hold phrase (with a repeat interval and a maximum number of repetitions).
Skills can override the tool-call filler phrases per tool and per language. See Advanced skill settings.
Save your changes
Click Save in the header to persist your voice configuration into a draft version.
Common issues
- The voice list is empty or limited — Some voices and languages are only available on certain TTS models. Switch the TTS model.
- The agent talks over the caller / can’t be interrupted — Enable Barge-In.
- Calls end too early — Increase Silence Timeout and/or Max Call Duration.
- Native Realtime is unavailable — It’s a Beta pipeline and may be gated by plan.