Pick and preview a voice, decide who speaks first, steer tone, and teach the agent to say your product names correctly.
Everything on the Voice tab that governs how the agent sounds: the voice itself, the first thing it says, tone direction, pronunciation, and the room it appears to be sitting in. Turn-taking and the silence timers are on the same tab but are a subject of their own — see Turn-Taking, Interruptions and Dead Air.
agents:voice:manage permission.Every row shows the voice's language, gender and style on one line and, in small type beneath, the provider and model that speaks it:
| Family | Count | Small label | Notes |
|---|---|---|---|
| Cartesia | 42 | Cartesia · Sonic 3.6 |
English (incl. EN-US, EN-GB), Spanish, French, Hindi and Punjabi voices, plus ten tuned for emotional range (Emotive) |
| OpenAI | 10 | OpenAI · gpt-4o-mini-tts |
Alloy, Ash, Ballad, Coral, Echo, Fable, Onyx, Nova, Sage, Shimmer — take a written Speaking style |
| Google (beta) | 30 | Google · Gemini 3.8 Flash-Lite TTS (Fast) or Google · Gemini 3.8 Flash TTS (Expressive) |
Google's named voices (Kore, Puck, Achird, Sulafat, …), each described in a word (Firm, Upbeat, Friendly, Warm). Every one speaks every language, and they take a Speaking style. A super admin enables Gemini voices per workspace while they are in beta; until then they are not listed, and an agent cannot be saved with one |
| Google library (beta) | Thousands | Same as Google | Every other voice in Google's catalog, marked Library: voices by language and regional accent (Canadian French, Glasgow English, Egyptian Arabic, …), each with a gender, a pitch and a persona such as Receptionist or Tutor. The list comes from Google, so new voices appear as Google adds them. They take a Speaking style and both voice qualities. Not on Gemini Live (see below) |
Your voices appears first when your workspace has custom voices: a person on your team cloned with their consent, or a voice designed from a description. Each is marked Yours. Calls in one add $0.06 a minute, and an agent on one goes live only when its opening says it is an AI and that the call may be recorded — see Custom Voices.
If the agent already holds a voice id that is not in the grid, the picker says so underneath: "Custom voice ID in use: …". It still works; it simply is not one of the listed options. When the id is one of your custom voices that can no longer speak (deleted, or custom voices switched off), the picker names the voice and says calls speak a standard voice instead.
The choice decides which speech engine runs the call — the small label on the card names it. That is why the next settings only appear for some voices.
When the agent speaks with a Gemini voice (its own voice, or a language's voice under Multi-Lingual Support), a Gemini voice quality choice appears under the picker:
| Choice | Model | What it is for |
|---|---|---|
| Fast (the default) | Gemini 3.8 Flash-Lite TTS | Starts speaking quickest. Every Gemini voice used this before the choice existed. |
| Expressive | Gemini 3.8 Flash TTS | Richer acting, with laughs, sighs and pauses that land. A little slower to start speaking. |
The preview buttons play the quality you pick, and every Gemini voice the
agent uses follows it, including your own Gemini voices and the short lines it
says while the caller waits. The choice does nothing for Cartesia or OpenAI
voices or on a live-audio model, and it is kept if you move away from Gemini,
so it applies again if you come back. Over the Karmaflow.ai connector,
create_voice_agent and update_voice_agent take it as geminiTtsQuality
(fast or expressive).
Speaking style, under the voice picker on the Voice tab, is a short line on how the voice delivers every sentence — "warm and unhurried", "brisk and factual". The preview buttons use it, so you hear the style before you save. Under the field, Try: offers a few examples; tap one to fill the field (and Undo to get your own text back), then adjust it. On a Gemini Live model the examples include accents; on other voices they don't, because a Gemini voice's accent comes from the voice you pick. For a regional accent, pick it from Google library (search for the accent) rather than writing it into the style: Google names accents in a style as a cause of a voice drifting mid-call.
Left empty, the field uses Karmaflow.ai's default style, and the form shows it under the field: a short, warm line for OpenAI and Gemini voices, and a fuller one for Gemini Live agents, which can judge from the caller when a soft chuckle or a thoughtful "hmm" fits (unless the platform's default has been turned off, in which case the form says nothing). Anything you type replaces the default for that agent; a style over 300 characters is not used at all, and the form says so.
It appears for OpenAI and Gemini voices, the two engines that take a written style, and on a Gemini Live model (below). Keep it short: a long description makes a Gemini voice drift, and the field warns you past about 120 characters. Cartesia voices take their tone from the words themselves, so the field is hidden for them; if you switch an agent with a style to a Cartesia voice, the text is kept and applies again if you switch back. To change a Cartesia voice's feel, change the voice.
Tone that belongs in the instructions instead: how long turns should be, whether to use the caller's first name, how formal to be. Those are behavioural, and they work on every voice.
Who speaks first sits with the greeting it controls, in the Opening group at the top of Conversation → Instructions & Greeting (moved from this tab on 2026-11-02). Two values:
The greeting is right under it, with Hear it to play it in the agent's
voice. It is spoken verbatim, so it is the one line of the call you
control exactly. {{first_name}}, {{last_name}} and {{caller_number}}
resolve in it; an unknown caller resolves them to empty strings, so
Hi {{first_name}}, thanks for calling becomes Hi , thanks for calling —
write it as Hi, thanks for calling and let the instructions handle names.
Outbound calls behave differently on purpose: the agent waits for the callee to say something (up to 30 seconds) before it plays its opener, and plays it without allowing an interruption. Picking up a call and being talked over is the thing that makes an outbound robot obvious.
Voice delivery tags in the greeting are deliberate. The static greeting is the one place emotion markup may be authored inline, and it appears as written in the stored transcript. That is expected. Anywhere else in a call, the platform rules forbid the agent emitting those tags and they are stripped — the browser test panel flags such a turn with a small tags marker so you can report it.
Voicemail is set on the Call handling tab, in the Voicemail card. It decides what the agent does when an outbound call, or a warm transfer, reaches an answering machine: leave a message in its own words (default, guided by your Voicemail instructions), say an exact message, or hang up without a message. Detection hints, under Help it recognize a machine, help the platform recognise your customers' machines. Details and the full table are in Placing Outbound Calls.
Paused on calls (October 2026). Agents currently say every word as written while we test which entries the latest voices still need. Your packs and workspace entries are kept and apply again when pronunciation is switched back on.
Pronunciation Packs (below the voice picker) rewrite words on their way to the speech engine, so your product and place names come out right instead of phonetically plausible. Packs come in four groups:
| Group | Behaviour |
|---|---|
| Always-On (Platform Core) | Applied to every agent; not switchable |
| Regional Packs | Opt in per agent — US, UK, Canadian English and French, French, Spanish and Mexican Spanish |
| Vertical Packs | Opt in per agent — automotive, finance, healthcare, home services, hospitality, legal, real estate |
| Tenant Custom | Your workspace's own entries; always applied |
Each group is closed until you click it; its heading says what is on ("2 on: US English, Healthcare" or "None on") and how many packs it holds. Open a group and tick the packs this agent should carry. Each row shows the pack name, what it covers and an entry count.
A word still coming out wrong after the right pack is on usually needs a workspace entry rather than a pack — that is what Tenant Custom is for.
Background Audio (default on) plays quiet office ambience and keyboard typing while the agent is thinking. It exists because dead silence during a pause reads as a dropped call, and a caller who thinks the line dropped hangs up.
It only plays on telephony calls. Browser calls never get it, so on a web agent the switch is replaced by a line that says so.
Turn it off for agents where the illusion is wrong — an explicitly automated line, or a clinical or legal context where ambience is out of place.
If a super admin has enabled Gemini Live voice agents (beta) for your workspace, the AI Model list on the Setup tab also offers Gemini 3.8 Live and its extended thinking variant. On one of these a single Google model listens and speaks: there is no separate speech-to-text or text-to-speech step, and the Voice tab changes to match.
{{first_name}} placeholder resolved empty
for an unknown caller. Remove it from the greeting.