Voice, language and behavior settings

The Voice, language & behavior tab of the agent editor is the full tuning surface for how your agent sounds and behaves on a call. Change what you need, click Save settings, and the configuration syncs to the voice engine.

Voice

The voice list is dynamic: it loads live from your Retell account, so it shows exactly the voices your account can use — including ElevenLabs voices if you bring your own ElevenLabs key on Retell. Filter by name, provider or accent, and use Preview voice to listen before committing: it plays a three-second sample — enough to judge tone and accent — and stops on its own. If the list cannot load (for example Retell is not connected yet), you can still paste a voice ID manually.

Below the picker you can tune the voice itself: Voice model (TTS), Speed, Temperature (expressiveness) and Volume. The voice model list follows Retell's catalog — for ElevenLabs voices it includes ElevenLabs v4 Turbo, the real-time model with the lowest latency, and for Inworld voices Inworld TTS 2 and Inworld TTS 2 Flash; leave it empty to keep Retell's default for that voice.

Expressive mode appears in the same section only when the selected voice is one of Retell's platform voices; with ElevenLabs, Cartesia, MiniMax or other providers a short note tells you it is not available. Turn it on and the agent adds expressive tags to its own speech — an empathetic tone, a sigh, a pause, emphasis. The tags you tick (emotions, effects, emphasis) are a palette, not a script: the agent decides when to use them, and with none ticked it follows Retell's general expressive guidance. Tags you write by hand in square brackets in the prompt are still allowed.

This works for every setting in the tab: emptying a field or picking Default (Retell) and saving also puts the Retell agent back on its default. Keys you remove from the Advanced config (JSON) are the exception: they are not reset on Retell.

LLM model and language

The LLM Model & Language section sets the brain and the tongue: pick the LLM model the agent thinks with, and the Language it speaks and understands (default: US English).

When Retell retires a model, GodSetter automatically shows and uses the model Retell names as its replacement, and the retired model disappears from the menu: an agent set to gpt-4.1-nano, for example, now runs on gpt-5.6-luna.

At the bottom of the menu, the Realtime (speech-to-speech) group lists the OpenAI realtime models (gpt-realtime-2.1, gpt-realtime-2.1-mini…): the model hears and speaks by itself, so the agent answers faster than with a text model plus a separate voice, but it speaks only with OpenAI voices (the voice list narrows to them) and costs more per minute (from $0.07/min for the mini models to $0.38/min, voice included).

Cost per minute and latency preview

Next to the model picker every model shows Retell's $/min, and under it a live estimate line updates as you change voice, model, noise reduction or the Advanced JSON: the per-minute cost (LLM + voice + Retell voice infrastructure, plus paid add-ons such as knowledge base, advanced denoising, PII removal or guardrails) and the expected end-to-end latency (voice + model + transcription). The same figures sit in the agent header. No test call is needed: the estimate comes from Retell's published price list, the same table Retell's own dashboard uses. Twilio telephony is not included — it is billed by Twilio on your trunk.

Two things to know:

• A * next to the cost means one component has no published price yet: the total is a floor, not the full figure.
• Once the agent has real calls, a second line shows the measured figures: the real spend per conversation minute (total cost ÷ total minutes) and the latency as p50 / p90, over the last real calls — test calls are excluded. That line is the truth; the estimate is the forecast.

Behavior

The Behavior section is a set of sliders and switches:

Audio, timing, transcription, privacy

Four more sections cover the rest of the call experience: Audio environment (ambient sound and its volume, noise reduction), Timing & timeouts (first message delay, end call after silence, reminders, max call duration, ring duration), Transcription (STT mode, specialized vocabulary) and Privacy & data (data retention).

A few of these menu choices are worth knowing before you pick them:

• Noise reduction: Noise + background voices also reduces the voices of other people near the lead. It is a paid add-on and shows up in the cost estimate as Advanced denoising. It can also filter out a lead who is far from the microphone or speaks quietly, so try it before using it on real leads. Recordings keep the original audio.
• STT mode: Custom (JSON) needs a custom_stt_config in the Advanced config (JSON) field described below.
• Specialized vocabulary applies only to agents that speak English; in any other language it has no effect.
• Data retention covers what Retell stores, not what GodSetter receives and archives. All data includes transcripts, recordings and logs, and it is Retell's default. Basic attributes only leaves them out. All data except PII removes only the personal data of the categories listed in pii_config (Advanced config (JSON)), when detected: with no categories, Retell stores everything.

In Timing & timeouts every duration is typed in seconds, decimals allowed, and both separators work — 2,5 and 2.5 are the same value. (They used to be milliseconds: nothing stored changed, only what you type.) Each field has its own range:

Max reminder count is the exception: it stays a plain number — how many reminders, not how long. A value outside its range shows the error under the field and blocks the save, so a typo never reaches the voice engine.

What the agent says to open the call — and whether it speaks first at all — is not configured here: it lives in the Call opening card at the top of the Prompt tab, see Create your first AI agent.

Advanced (JSON)

Nested configurations that don't have a dedicated field — pronunciation dictionaries, custom transcription config, PII handling and similar — go through the Advanced config (JSON) escape hatch. Paste a JSON object; it is merged with the fields above at save time. It must be a valid JSON object, and the form tells you if it isn't.

The same Advanced section carries a Time zone (IANA) field: it is the voice engine's own agent time zone, unrelated to the dialer's calling window — that one lives in Retries & transfer, see Set a calling window per agent.

Voicemail: the default policy

Voicemail detection is always on — the agent never talks to an answering machine. The default policy is to hang up silently when a machine answers. If you'd rather leave a fixed message, that is configured per agent in the Retries & transfer tab — see Retry cadence, callbacks and human transfer.

On an inbound agent that tab shows only the transfer and appointment cards: calling window, redials and voicemail belong to the outbound agents whose call-backs it answers.

When changes take effect

Click Save settings and the new configuration applies from the next call — calls already in progress are not touched. The sync to the voice engine is best-effort: a momentary sync error is logged and never blocks the save.

Agents created before a platform feature shipped (for example new in-call tools) do not pick those additions up from a voice or LLM save. Use the Sync GHL tools button in the Data & GHL tab, or save the policy in the Retries & transfer tab: those two are what re-register the in-call tools, the post-call analysis and the voicemail policy on the voice engine.

Next steps

• Test your agent from the browser — hear the new voice and behavior right away.
• Prompt versions, A/B metrics and rollback — tuning the words, not just the voice.