# Changing the voice

Why a voice change is not a small edit, what the console tells you before you commit, and the weekly limit that exists because of it.

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

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

Changing a voice looks like picking a different name from a list. It is closer
to re-recording every fixed line your agent owns.

## Why it costs something

Every phrase you have authored is recorded in a specific voice. Change the
voice and none of those recordings are usable. They are not wrong, they are
someone else, and playing them in a call that is otherwise in the new voice
would be worse than not having them at all.

<Figure
  label="A single voice change invalidating every recorded phrase"
  caption="The count is per language. An agent speaking two languages with seven phrases each has fourteen recordings to remake, and only the language whose voice you changed is affected."
>
<svg viewBox="0 0 700 180" xmlns="http://www.w3.org/2000/svg">
  <text x="0" y="14" class="d-eyebrow">ONE VOICE CHANGE, EVERY PHRASE</text>

  <rect x="0" y="52" width="132" height="52" rx="8" class="d-box-accent d-pulse" />
  <text x="14" y="74" class="d-label">New voice</text>
  <text x="14" y="90" class="d-sub">one setting</text>

  <path d="M136 78 L154 78" class="d-arrow" />
  <polygon points="160,78 153,74.5 153,81.5" class="d-arrow-head" />

  <rect x="172" y="30" width="120" height="38" rx="8" class="d-box" />
  <text x="184" y="54" class="d-sub">Greeting</text>
  <rect x="302" y="30" width="120" height="38" rx="8" class="d-box" />
  <text x="314" y="54" class="d-sub">Acknowledgement</text>
  <rect x="432" y="30" width="120" height="38" rx="8" class="d-box" />
  <text x="444" y="54" class="d-sub">Acknowledgement</text>
  <rect x="562" y="30" width="120" height="38" rx="8" class="d-box" />
  <text x="574" y="54" class="d-sub">Acknowledgement</text>

  <rect x="172" y="88" width="120" height="38" rx="8" class="d-box" />
  <text x="184" y="112" class="d-sub">Idle nudge</text>
  <rect x="302" y="88" width="120" height="38" rx="8" class="d-box" />
  <text x="314" y="112" class="d-sub">Idle nudge</text>
  <rect x="432" y="88" width="120" height="38" rx="8" class="d-box" />
  <text x="444" y="112" class="d-sub">Farewell</text>
  <rect x="562" y="88" width="120" height="38" rx="8" class="d-box" />
  <text x="574" y="112" class="d-sub">Fixed answer</text>

  <text x="172" y="152" class="d-accent-text">every one of these has to be recorded again before a caller needs it</text>
</svg>
</Figure>

## What the console shows before you commit

Picking a voice does not change anything. The console first shows you what the
change would cost:

- **How many phrases** are affected, in that language.
- **How much audio** that is in total, as the caller experiences it.
- **How long re-recording takes**, which is the number that decides whether a
  caller is going to wait.
- **Whether those figures are measured or estimated.** Before a phrase has ever
  been recorded, its length is calculated from the text. The console says so
  rather than presenting a guess as a measurement.

You confirm against those numbers. Nothing moves until you do.

## The order it happens in

The new recordings are made **before** the change goes live. Until they exist,
calls keep using the previous voice.

This is the part worth understanding, because the obvious order is wrong.
Switching the voice first and recording afterwards produces a window where some
phrases exist in the new voice and some in the old, and a caller in that window
hears two different people in one conversation. Recording first means the worst
case is a call that sounds exactly like yesterday.

<Aside type="caution" title="If a call arrives before recording finishes">
That caller waits for whichever phrase is needed first, once. It is the same
wait that existed before any recording was kept at all, and it applies to one
call rather than every call. After that the phrase is ready.
</Aside>

## The weekly limit

You can change a voice **three times per language, per week**. An owner can
override the limit when there is a real reason to.

The limit is not about cost. A voice is what your business sounds like, and a
business that sounds different on Tuesday than it did on Monday reads as
unreliable to the people who call it more than once. Three is enough to
audition alternatives and settle. It is not enough to keep fiddling.

When the allowance is used up, the console says how many changes remain and
when the window resets.

## Every change is recorded

Voice changes are logged: which language, from which voice to which, by whom,
and when. Two reasons this matters. Callers report that "it sounds different"
long after anyone remembers making a change, and a voice that turned out to
mispronounce something specific is worth being able to trace back.

## Practical advice

- **Audition before you commit.** Listen to a real phrase of yours in the
  candidate voice, not a vendor sample of somebody else's sentence. Your
  business name is usually where a voice fails.
- **Change one language at a time.** Two changes at once means two sets of
  recordings and no way to tell which one people are reacting to.
- **Do it when the line is quiet.** Not because anything breaks, but because
  the one caller who might wait should be nobody.
- **Re-read your phrases afterwards.** A line written for a brisk voice can
  sound curt in a warm one. The words are half of the impression.
