Voice & conversation
On this page
Two people report the same complaint about voice AI more than any other: “it talks over me” and “it takes forever to answer”. Both are on this page, and both are one slider.
Everything here lives at AgentsEditVoice.
Choosing a voice
Ten voices ship with every account. Play the samples and pick — this genuinely is a five-second decision, and you can change it whenever you like.
| Voice | Character | Suits |
|---|---|---|
| Alloy | Neutral, even | Support lines, anything you want to feel unremarkable |
| Nova | Bright, warm | Reception, retail, appointment booking |
| Shimmer | Soft, calm | Healthcare, sensitive or difficult conversations |
| Coral | Friendly, upbeat | Sales, outbound, hospitality |
| Sage | Measured, steady | Finance, insurance, anything that needs to sound solid |
| Echo | Clear, straightforward | General purpose, logistics, dispatch |
| Ash / Ballad / Verse / Marin | Alternatives across the same range | When the obvious pick is not quite right for your brand |
Phone audio is narrow and compressed, and voices that sound identical through good speakers separate noticeably down a line. Use Test Call and listen on the handset your callers will use.
Turn-taking: the four timers
The hardest problem in voice AI is knowing when the caller has actually finished speaking. A person pausing to think and a person who has finished sound identical for the first half-second. These four settings are how you tell the agent which way to lean.
| Setting | Default | What it means | Raise it when… |
|---|---|---|---|
| Wait seconds | 0.5s | The base pause after the caller stops before the agent begins. | The agent cuts people off generally. |
| On punctuation | 0.2s | The wait when the caller clearly finished a sentence. | Rarely. Low is right — a finished sentence deserves a fast reply. |
| On no punctuation | 1.5s | The wait when the caller trailed off mid-thought. | Callers are elderly, unwell, thinking aloud, or reading from a document. |
| On number | 1.0s | The wait when the caller is in the middle of digits. | You collect phone numbers, account numbers or dates of birth. |
There is no setting that gives you both. Every extra tenth of a second you add to stop interruptions is a tenth of a second of silence on every single turn. Tune to the least-bad answer for your callers, not to a theoretical ideal.
How to tune in practice
- Change one timer, by 0.2–0.3s. Not all four.
- Make a real test call and deliberately trigger the behaviour you are fixing.
- Read the transcript in Call History. A cut-off caller shows up as a suspiciously short line ending mid-sentence — clearer than trying to remember how it felt.
- Repeat until it is acceptable, then leave it alone.
Settings for common situations
Starting points, not laws. Each is a small edit away from the default.
| Situation | Wait | No punctuation | On number | Why |
|---|---|---|---|---|
| Standard reception | 0.5s | 1.5s | 1.0s | The defaults. Leave them until something annoys you. |
| Collecting numbers or IDs | 0.6s | 1.6s | 1.6s | People read digits in bursts. Cutting in halfway through an account number is fatal. |
| Elderly or healthcare callers | 0.9s | 2.2s | 1.6s | Patience beats speed. Being interrupted is far more upsetting than waiting. |
| Fast B2B qualification | 0.4s | 1.2s | 1.0s | Busy people answering in short bursts want it snappy. |
| Noisy environment (street, shop floor) | 0.7s | 1.8s | 1.2s | Background noise makes the end of speech harder to detect. Give it room. |
Noise and background audio
Noise suppression
Cleans the caller’s incoming audio before the agent tries to understand it. Options are Off, Krisp and DeepFilterNet.
- Off — the default, and correct when your callers are mostly indoors on decent connections.
- Krisp — the general-purpose choice. Turn it on if your callers phone from streets, markets, cars or factory floors.
- DeepFilterNet — an alternative worth trying if Krisp is not handling your particular noise well.
Filtering adds a little processing time to every turn, and aggressive filtering can clip quiet speech. Turn it on because you have a noise problem, not as a precaution.
Background audio
Optionally plays faint ambience behind the agent — Office, Café or Nature. Total silence behind a voice sounds slightly uncanny, and a low office hum makes some agents feel more natural.
It is a matter of taste and it is not always an improvement — test it both ways and keep whichever you prefer. Leave it off for anything formal.
Engines
Which underlying speech services do the listening and the speaking. Both default to Vociply, which is tuned for phone audio and is the right answer unless you have a specific reason.
| Options | When to change | |
|---|---|---|
| Voice engine (speaking) | Vociply (default), AWS Polly, Azure Speech, ElevenLabs | ElevenLabs for a specific branded or cloned voice. Polly or Azure if you already standardise on them. |
| Speech engine (listening) | Vociply (default), Deepgram, Azure Speech, Google Speech | If a particular accent or dialect in your market transcribes poorly, trying another engine is the fastest fix. |
When an agent answers something the caller never asked, check what the transcript says it heard. Heavy accents, poor lines and unusual names are often mis-transcribed, and the agent is answering the question it received. That is a speech-engine problem, not a prompt problem.
Voice questions
Can I clone my own voice, or a specific person’s?
Through the ElevenLabs engine, yes. You need consent from whoever the voice belongs to — in many places using someone’s voice without it is illegal, and it is a fast way to lose customer trust.
The agent mispronounces our company name.
Spell it phonetically in the instructions: Our name is Njiru — say it “n-JEE-roo”. This works for product names, place names and surnames too.
Can it switch language mid-call?
It will generally follow a caller who switches, but reliability varies by language pair. If you serve two languages seriously, run two agents on two numbers — the quality is noticeably better than one agent trying to do both.
Why does it pause before answering questions about my calendar?
Because it is actually checking. A tool call adds a second or two while it queries the real system. You can cover it in the instructions: “If you need to check something, say ‘let me look at that for you’ first.”
Callers keep hanging up in the first few seconds.
Almost always the greeting. Long, corporate or robotic openings lose people. Short, human, and immediately useful — “Brightsmile Dental, how can I help?” — keeps them. See Writing instructions.