# Spoken phrases

Greetings, acknowledgements, idle nudges, farewells and fixed answers: the five kinds of phrase a call can speak, and how to write each one.

Source: https://docs.omazy.ai/how-to/voice/phrases/

import Figure from '../../../../components/Figure.astro'
import { Aside } from '@astrojs/starlight/components'

A voice agent writes most of what it says on the spot. A handful of lines
should never be improvised: the first sentence a caller hears, the noise it
makes while thinking, and the way it ends a call.

Those are spoken phrases. You author them under **Agent, Saved Responses** on
the **Voice** tab, one version per language, and the call speaks them word for
word.

They are the same objects as [saved responses](/how-to/agent/saved-responses/) in
chat. A saved response becomes speakable when you give it a spoken role.

## The five kinds

<Figure
  label="When each kind of phrase is spoken during a call"
  caption="Only the greeting is guaranteed to play on every call. The rest fire on a condition, which is why a line that is missing one of them fails in a way that is easy to miss."
>
<svg viewBox="0 0 700 200" xmlns="http://www.w3.org/2000/svg">
  <text x="0" y="14" class="d-eyebrow">WHEN EACH KIND IS SPOKEN</text>

  <line x1="0" y1="110" x2="698" y2="110" class="d-arrow" />
  <polygon points="698,110 691,106.5 691,113.5" class="d-arrow-head" />

  <text x="70" y="74" class="d-label" text-anchor="middle">Greeting</text>
  <text x="70" y="90" class="d-sub" text-anchor="middle">line connects</text>
  <path d="M70 96 L70 106" class="d-arrow" />
  <circle cx="70" cy="110" r="4" class="d-box-accent" />

  <circle cx="215" cy="110" r="4" class="d-box" />
  <path d="M215 114 L215 124" class="d-arrow" />
  <text x="215" y="140" class="d-label" text-anchor="middle">Acknowledgement</text>
  <text x="215" y="156" class="d-sub" text-anchor="middle">agent is thinking</text>

  <text x="360" y="74" class="d-label" text-anchor="middle">Answer</text>
  <text x="360" y="90" class="d-sub" text-anchor="middle">trigger matched</text>
  <path d="M360 96 L360 106" class="d-arrow" />
  <circle cx="360" cy="110" r="4" class="d-box" />

  <circle cx="505" cy="110" r="4" class="d-box" />
  <path d="M505 114 L505 124" class="d-arrow" />
  <text x="505" y="140" class="d-label" text-anchor="middle">Idle nudge</text>
  <text x="505" y="156" class="d-sub" text-anchor="middle">caller goes quiet</text>

  <text x="645" y="74" class="d-label" text-anchor="middle">Farewell</text>
  <text x="645" y="90" class="d-sub" text-anchor="middle">call ends</text>
  <path d="M645 96 L645 106" class="d-arrow" />
  <circle cx="645" cy="110" r="4" class="d-box" />
</svg>
</Figure>

| Kind | Spoken when | How many to write |
|---|---|---|
| **Greeting** | The call connects | One per language |
| **Acknowledgement** | The agent is working on a reply | Three or four per language |
| **Idle nudge** | The caller has gone quiet | Two or three per language |
| **Farewell** | The call is ending | One per language |
| **Answer** | A trigger phrase matches | As many as you have fixed answers |

<Aside type="caution" title="Only published phrases reach a call">
A phrase left in draft is invisible to the phone. This catches people out most
often with a greeting: the words are written, the voice is chosen, and the call
still opens with the platform default because nobody pressed publish.
</Aside>

## Writing a greeting

The greeting is doing more work than any other line. It is the caller's entire
first impression, it happens before they have said anything, and it sets what
they think they are allowed to ask for.

A greeting that works has four parts, in this order:

1. **Who is speaking.** A name. Callers talk more naturally to something named.
2. **Where they have reached.** The business, said plainly.
3. **What this line can do.** One clause. This is the part that stops the
   caller opening with "is this a real person".
4. **A question.** Hand the turn over explicitly, or you get silence.

> Good afternoon, you have reached Weston Dental on Queens Boulevard. I can
> book appointments, answer questions about treatments and fees, and take a
> message for the practice. How can I help you today?

Two things to avoid. Do not stack more than one sentence before the question:
the caller is waiting to speak, and a long opener is heard as a recorded menu.
And be deliberate about whether you announce what the agent is. Some practices
want it said outright, some want the caller to simply be helped. That is a
positioning decision, not a technical one, and it belongs to you.

<Aside type="tip" title="Length is a real constraint">
The console shows the spoken length of each phrase as you write it. A greeting
over about ten seconds is long enough that callers start talking over it. If
yours is long, cut the description of what the business does before you cut the
question.
</Aside>

## Writing acknowledgements

An acknowledgement is what the line says while the agent is working out an
answer. On a phone call, two seconds of silence reads as a dropped connection.

Write three or four per language and keep every one of them short:

> One moment.
>
> Let me check that for you.
>
> Sure, just looking that up now.

They are used in rotation. That matters more than it sounds: hearing the
identical sentence twice in one call is the single clearest signal that there
is a machine on the line.

Avoid anything that promises a particular kind of answer. A phrase like "I will
answer briefly" commits the agent to something it may not do, and a caller
notices the mismatch immediately.

## Writing idle nudges

An idle nudge re-engages a caller who has stopped talking. Each one is spoken
followed by the question the agent last asked, so the caller hears what they
were actually asked rather than a bare "are you there".

> Sorry, are you still there?
>
> Take your time.

Two or three, again in rotation, again short.

## Writing a farewell

One line, per language. This is the last thing a caller hears, so unlike the
other rotating kinds there is exactly one and it is always the same. A call
that ends differently depending on which phrase came up would be strange in a
way people notice without being able to say why.

## Fixed answers

An answer phrase is spoken instead of a model-written reply when a trigger
phrase matches. Use it for anything that must go out in exact wording: opening
hours, a regulated disclosure, a price that must not be paraphrased.

Everything else should stay model-written. A voice agent whose answers are all
fixed is a phone menu with extra steps, and callers treat it like one.

## Languages

Every phrase is authored per language. A missing language is skipped rather
than substituted, which is the safe behaviour: hearing a Bengali greeting
followed by an English acknowledgement is worse than hearing neither.

If you add a language after the fact, walk back through every kind on this page
and write its version. The console shows you which are missing.
