Voice and fillers
How the agent sounds and behaves on a call, from voice, speed, tone and language to the greeting, turn-taking, ambience and filler phrases.
Two pages decide how the agent sounds out loud and how it carries a conversation: the Voice page picks the voice and the language it listens in, and the Behaviour page tunes the greeting, speaking style, turn-taking, ambience, and filler phrases. Together they are the difference between a call that feels human and one that feels robotic, so they earn the same care as the prompt and tools.
Every setting on both pages lands in a draft and only reaches live calls once you publish. One of them, the voice, must be set before you can publish at all (see below).
Voice
The Voice page sets the Cartesia voice the caller hears, plus the speed, tone, and the language the agent listens in.
Pick and hear a voice
Browse the voice list and press play on any one to hear it before you commit. Picking a voice you have actually listened to is the point: a voice that reads well on paper can land wrong on a phone speaker.
A brand-new agent ships with no voice selected on purpose. That is the single blocking error a fresh agent carries: the publish gate refuses with "No voice is selected" until you choose one, so you never ship a voice nobody heard. If a call somehow runs before a voice is set, the runtime falls back to the fleet default, Thandi - Direct Dispatcher, so a live call never breaks for the want of a voice. That fallback is a safety net, not a substitute for choosing.
Thandi is the fleet-wide default, used as the preview voice and the runtime fallback. It is a good neutral starting point, but pick the voice that fits the business rather than leaving the default in place.
Speed
Playback rate from 0.5× to 2.0×, in 0.1× steps. Around natural pace is usually best: too fast reads as rushed, too slow as robotic. Use the Preview on the Behaviour page (Speech category) to hear the current speed before you save.
Tone
An emotional colour for the voice. On a Cartesia voice this is the emotion control, with eleven presets:
| Value | Reads as |
|---|---|
| Neutral | Even, no colour (a safe default) |
| Content | Settled, at ease |
| Happy | Bright, upbeat |
| Excited | High energy |
| Calm | Slow, reassuring |
| Confident | Assured, in control |
| Sympathetic | Warm, understanding |
| Curious | Engaged, inquisitive |
| Sad | Subdued |
| Angry | Sharp (rarely useful for an inbound line) |
| Scared | Tense (rarely useful for an inbound line) |
Match the tone to the job: a support line wants calm or sympathetic, a sales line might want confident or content. A freshly created agent starts on a medium positive tone.
Language
The language the agent listens and speaks in. Eleven options:
| Value | Language |
|---|---|
| en | English |
| es | Spanish |
| fr | French |
| de | German |
| it | Italian |
| pt | Portuguese |
| ja | Japanese |
| ko | Korean |
| zh | Chinese |
| hi | Hindi |
| ar | Arabic |
This sets both the speech-to-text language the agent hears and the voice it speaks back in, so set it to match your callers.
The greeting
The greeting is the first thing a caller hears when the agent picks up. It is spoken word for word the instant the call connects, straight to the voice, skipping the model round-trip. That is what makes the open of a call instant and consistent rather than starting with a short pause while the model thinks.
Every new agent ships with a greeting already set, built from the workspace name, for example:
Thanks for calling Mitchell's Plumbing. How can I help you today?
If the workspace has no name yet, it falls back to a plain "Thanks for calling. How can I help you today?" Edit it to whatever fits, or turn the fixed greeting off entirely. With it off, the runtime falls back to a greeting the model generates on the spot, which costs you that instant first speech, so a fixed greeting is the better default.
A good greeting is short and tells the caller who they have reached. Naming the business in the first line sets the tone for the whole call.
Pronunciation and recognition hints
Two settings help the agent hear and say tricky words. Both live on the Behaviour page and both are agent-wide (they apply across every version, not just the draft).
Recognition hints (keyterms)
The agent's ear for tricky words. Recognition hints are words and phrases the speech-to-text engine is told to listen out for, so it does not garble a brand name, a suburb, a product code, or a person's name. The runtime passes them to Deepgram as keyterms (up to 100 per agent), which boosts how reliably those exact terms come back in the transcript.
Add the proper nouns your callers actually say: your business name, your product range, the suburbs you serve. If "Joondalup" keeps coming back as gibberish, it belongs here.
Pronunciation
The agent's tongue for tricky words. A pronunciation entry teaches the voice how to say a word it would otherwise mangle: a brand name, a place, an acronym, a bit of jargon. Add the word and how it should sound, and the runtime compiles the list into a Cartesia pronunciation dictionary for the voice.
Recognition hints fix what the agent hears; pronunciation fixes what it says. A name like a clinic or a product range usually wants both.
Conversation and behaviour
The rest of the Behaviour page tunes how the agent carries the conversation: turn-taking, response length, ambience, and caller memory.
Turn-taking
How patiently the agent waits before it decides the caller has finished and replies. This is the difference between an agent that talks over people and one that leaves awkward gaps.
- Turn-start detection decides which signals open a caller's turn. Standard (recommended) picks up quiet or trailing openers and is the right choice for most lines. Minimal opens a turn only on enough words. Fast barge-in lets callers cut in quicker, which is great on a handset but risky on a speakerphone, where the agent can mistake its own echo for the caller.
- Smart interruption filter holds off replying when the caller pauses mid-thought, for example while reading out a code in fragments. It waits for the completed turn instead of jumping in on the first pause.
- Response sensitivity and Quick-response threshold trade certainty against speed: a higher response sensitivity waits for more certainty before replying (good for quiet, hesitant speech), while a higher quick-response threshold replies sooner in fast back-and-forth (lower latency).
- Short pause timeout and the long-turn ceiling set how long a silence runs before the agent steps in.
The defaults are tuned for a typical inbound phone call, so leave them alone unless you have a specific problem to fix, and change one knob at a time.
Response length
How long the agent's spoken replies run. The control is a slider from brief to thorough, with named anchors:
| Anchor | Reads as |
|---|---|
| Brief | Tight, transactional answers |
| Standard | A sensible default |
| Detailed | Fuller answers |
| Thorough | Long, complete answers |
On a phone call, shorter is almost always better: long monologues are hard to follow by ear and slow the back-and-forth.
Response length is a soft preference, not a hard cap. The runtime turns your setting into a plain instruction in the system prompt ("keep each spoken reply to about N words"), expressed as an approximate word count. It deliberately never lowers the model's own token ceiling, so a long tool call can never be silently cut off. It nudges the agent to be brief; it does not gag it mid-action. The model picker is on the Model page.
Office ambience
An optional low background audio bed, a quiet call-centre hum, so the agent does not sound like it is speaking from a vacuum. It is a simple on or off toggle with a volume, and it is off by default. A little ambience can make a line feel more human; too much is distracting, so keep it subtle.
Caller memory
Whether the agent remembers people across calls. With it off, every call is treated as a first call. With it on, the agent builds a profile per phone number and can recall a caller's name and details on a callback. When it is on, you also choose how it uses what it knows:
- Check first asks before using a remembered detail ("Am I speaking with Sam?"). This is the safest option, since phone numbers get shared and handed on.
- Greet by name welcomes a recognised caller back by name straight away.
- Use quietly uses what is remembered during the conversation without announcing it.
You also choose whether to personalise the very first line for a recognised caller, or keep the fixed greeting and only use memory once the conversation is under way.
Fillers
Fillers are the short, natural phrases the agent says while it works, "one moment", "just a sec", so the caller is not met with silence while a tool runs. They are the single biggest lever on perceived latency: a filler at the right moment makes a slow lookup feel responsive even though nothing got faster.
How fillers work
When the agent calls a tool, if the tool takes longer than the filler delay (0.5 to 5.0 seconds) the runtime plays a phrase to cover the gap. Two controls sit on the Behaviour page:
- Play filler phrases turns the whole behaviour on or off.
- Filler delay sets how long the agent waits before reaching for a filler, so it does not interrupt a tool that returns quickly.
Default versus per-tool fillers
There are two pools of phrases:
- Default phrases are the agent-wide pool, edited on the Behaviour page. A new agent ships with two: one moment and just a sec. The agent picks from these whenever a tool runs and has no phrases of its own.
- Per-tool fillers are phrases specific to one tool, for example "looking up your order now", edited inside that tool's editor on the Tools page. A tool's own phrases override the default pool for that tool.
On a freshly created agent the filler phrases are seeded ready to use, but each tool ships with Play fillers off. That is deliberate: the house prompt already tells the agent to say a short, natural lead-in before using a tool ("let me check that for you"), and that spoken lead-in is the intended cover. A canned "one moment" on top of it just doubles up. Turn a tool's Play fillers on per-tool when you want a canned phrase instead of the model's lead-in.
Writing good fillers
Keep them short, natural, and varied. A handful of two-or-three-word phrases the agent can rotate through reads far better than one phrase repeated every time. Match them to the moment: a generic "one moment" works anywhere, while a tool-specific "checking that order for you now" tells the caller exactly what is happening.
Why end call has no fillers
The end call tool never plays a filler, and that cannot be turned on. A tool that hangs up the line must never say "one moment" first, then go silent, so the runtime enforces fillers off for it. New agents are seeded with no filler phrases for end call at all, by design.