# Custom Voices

A custom voice gives your voice agents a voice no other workspace has: a
person on your team, cloned with their consent, or a new voice you describe in
a sentence. Once it exists, it appears in the voice picker under **Your
voices** like any other voice.

Custom voices sit behind two checks: your **business is verified** once, and
**each cloned voice carries the speaker's recorded consent**. The rest of this
article walks through both and then shows how to use the voice.

## What you'll need

- A **tenant admin** login. Applying for verification and creating, renaming
  and deleting voices are tenant-admin actions.
- The **`agents:voice:manage`** permission to put a voice on an agent.
- Proof that your business is registered: a certificate of incorporation,
  business registration, or similar official document. PDF, PNG, JPEG or
  WebP, up to 5 files, 10 MB each.
- For a cloned voice: the speaker, a quiet room and a microphone. You can also
  upload recordings made elsewhere.

Custom voices are for **standard voice agents**. Agents on a live-audio model
(Gemini Live) keep their built-in voices for now; the picker does not offer
custom voices to them.

## How it works

```mermaid
flowchart LR
  A[Apply with your<br/>registration documents] --> B{AI check}
  B -- genuine, confident --> C[Verified]
  B -- unclear --> D[Karmaflow reviews]
  D --> C
  B -- fake --> E[Not approved]
  C --> F[Clone or design<br/>a voice]
  F --> G[Pick it on a<br/>voice agent]
  G --> H[Go live: opening says<br/>AI + recorded]
```

## Step 1: verify your business

Open **Settings (the gear icon) → Channels → Custom Voices**
(`/settings/custom-voices` *(sign-in required)*). The page shows two steps at
the top, **Verify your business** and **Create your voices**; the second opens
once the first is done.

The application has four numbered sections:

1. **Your business** — the legal name, registration number and country, as
   they appear on your registration, plus your website if you have one.
2. **Proof of registration** — drop files on the box ("Drop files here, or
   click to choose"). Clear, complete pages; a photo of the original is fine
   if every word can be read.
3. **The rules for custom voices** — tick each rule. They are part of your
   agreement with Karmaflow.ai:
   - Only voices given with consent
   - No public figures, minors or impersonation
   - Callers are told it is an AI
   - Consent records, and withdrawal the same day
   - We follow the laws that apply to us
   - We follow Google's and Cartesia's policies
   - Karmaflow may suspend a voice
4. **Sign** — your name and role, and your name again as your signature.

Click **Submit for verification**. An AI model reads your documents straight
away, usually within a minute, while the page shows **Checking your
documents**:

| What the check finds | What happens |
|---|---|
| The documents look genuine, match the name and number you entered, and the model is confident | **Verified.** Step 2 opens. No person needs to look. |
| Something is unclear: a name that does not quite match, a page that cannot be read | **With Karmaflow for review.** The page lists what the check noticed. A Karmaflow.ai administrator decides, and you get an email. |
| The documents show clear signs of being fake | **Not approved.** The page says why. |

If Karmaflow.ai asks for more (**Karmaflow needs a little more**), the page
says what is needed; fix it and submit again. Every outcome is emailed to the
person who applied.

Karmaflow.ai can also verify a workspace without an application. The page
then says **Verified by Karmaflow** and shows **Accept the rules for custom
voices**: a tenant admin ticks each rule and signs once, and **Create a
voice** opens.

<!-- SCREENSHOT: voice-agents/custom-voices-apply.png — Demo tenant, /settings/custom-voices, status "Not set up": the step indicator, the "How it works" tiles and the four numbered sections of the application, with one document attached. -->

## Step 2: create a voice

When the workspace is verified, **Your voices** appears below the application,
with a meter showing how many voices you have of your limit (two per
workspace; Karmaflow.ai can raise it). Click **Create a voice** and choose
**Clone a voice** or **Design a voice**.

### Clone a voice

The dialog takes you through four steps.

1. **Voice and speaker**
   - **Name for this voice** — what you will see in the picker.
   - Choose the engine:

     | | Cartesia | Google Gemini |
     |---|---|---|
     | Languages | Speaks the one language it is recorded in | Speaks every language from one recording |
     | Speed | Starts speaking fastest on calls | — |
     | Consent | Checked by Karmaflow.ai (the words are transcribed and compared) | Google checks the consent recording itself |
     | Voice sample | 10 to 60 seconds | 10 to 30 seconds |

   - **Whose voice is it?** — the speaker's full name.
   - For Cartesia, **Language the voice will speak** and **Consent statement
     in** (English or French). For Google Gemini, **Language the speaker reads
     the consent statement in**, one of the 30 languages Google provides a
     statement for; the voice itself will speak every language.
2. **Consent** — the speaker reads the sentence on screen aloud, word for word,
   in their own voice. Click **Record the statement**, or **Upload a recording
   instead**. Then tick the box confirming the speaker agreed to have their
   voice cloned and knows they can withdraw that consent at any time. The
   recording must be 2 to 30 seconds.
3. **Voice sample** — the same speaker reads the text under **Read this, as if
   on a call**. Click **Record the voice sample**. Speak naturally, at the
   distance you would on a phone. The field tells you if the take is too
   short, too long, too quiet or clipping; **Record again** replaces it.
4. **Create** — check the summary and click **Create**.

If the consent recording does not match the statement, or Google refuses the
recordings, the dialog says so and takes you back to the step to fix.

**What is kept:** the consent recording, as evidence, for as long as the voice
exists. The voice sample is sent to the voice provider and not stored by
Karmaflow.ai.

<!-- SCREENSHOT: voice-agents/custom-voices-clone-consent.png — Demo tenant, Create a voice → Clone a voice, step "Consent": the statement, the recorder with a level meter mid-take, and the consent checkbox. -->

### Design a voice

Designing needs no recording and no speaker; it uses Google Gemini.

1. Name the voice and, under **Describe it**, say what it should sound like:
   age, warmth, pace, accent. The chips under the box are starting points.
   **Main language** is optional; the voice speaks every language either way.
2. Click **Design the voice**. A sample plays in the dialog.
3. **Keep this voice** to add it to your voices, **Try again** for another
   take (it replaces the draft), or **Discard**.

A draft you don't keep is deleted after an hour. **Back** keeps the draft: the
library shows it until you keep or discard it.

## Use a voice on an agent

1. Open a voice agent and click the **Voice** tab.
2. In the voice picker, choose the **Your voices** category. Your voices are
   listed first, each marked **Yours**, with a play button.
3. Click a voice to select it, and save the agent.

You can also choose a custom voice per language under the agent's language
voices; it is listed as **Your voice:** followed by its name. A custom voice on
a language row counts like the agent's own voice: the check below applies, and
calls that switch into it carry the custom-voice charge.

Under the picker, whenever the agent speaks a custom voice, the form checks the agent's opening. An agent on a custom
voice **goes live only when its opening tells callers they are speaking with
an AI and that the call may be recorded**, either in the static greeting or in
the instructions for the first turn. If the opening is missing either part,
the form says which, and pushing the agent live is refused with the same
reason. See [Testing, Drafts and Going Live](/help/voice-agents/testing-and-promotion).

## Manage your voices

Each voice card on **Settings → Channels → Custom Voices** shows:

- whether it was cloned or designed, the engine and the language;
- for a cloned voice, **Consent of** the speaker **on file**, how it was
  checked and when;
- the agents that use it, each linked;
- for a Google voice, how long Google keeps it: a year after it last spoke,
  renewed every time it speaks.

Rename a voice with the pencil icon. If Karmaflow.ai switches custom voices
off for your workspace, the voices stay listed here, marked **Switched off**,
so you can still see which agents use them and delete them. **Delete** is disabled while any agent,
live or draft, uses the voice: choose another voice on those agents first.
Deleting removes the voice at the provider and deletes the consent recording;
the record of who consented, and when, stays.

## Billing

Calls spoken in a custom voice add **$0.06 per minute** on top of the call's
own minutes. The line appears on your usage as **Custom voice minutes**. A call
is charged the extra only once a custom voice has actually spoken on it — the
agent's own voice or a language voice it switched to. A call that could not use
the custom voice (see below) and never spoke it is not charged the extra.

## When a call cannot use the voice

If a voice was deleted, expired at Google, or custom voices were switched off
for your workspace — or your workspace has reached a cap Karmaflow.ai set on
custom-voice minutes — the agent does not go silent: it speaks a standard voice
instead — a Google voice in place of a Google custom voice, so it keeps
speaking every language, or the default Cartesia voice in place of a Cartesia
clone. The same happens for a custom voice on a language row. You can see it
in three places:

- the **call detail** page shows a **Custom voice** card saying the call spoke
  a standard voice and why — or, when the provider refused the voice mid-call,
  the minute it switched;
- the voice picker on the agent names the voice and says it is not available;
- the voice's card in your library is marked (**Switched off**, **Expired at
  Google**, **Not available**). A Google voice that Google no longer has is
  marked **Expired at Google** the first time a call finds out, so later calls
  go straight to the standard voice.

If you switch custom voices back on — Karmaflow.ai approves your workspace
again, including after you apply again — the voices that were switched off
come back as they were.

## Troubleshooting

**The application stays on "Checking your documents".** The check normally
finishes within a minute. If it does not, Karmaflow.ai reviews the application
by hand and you receive an email with the outcome.

**"The rules have changed".** Karmaflow.ai updated the rules. Voices you
already have keep working; tick and sign the new version to create another.

**The consent recording was refused.** Read the statement word for word, in the
language selected, in one take, with no one else speaking. For Google voices,
the consent and the sample must be the same speaker.

**"Create a voice" is disabled.** You are at your limit. Delete a voice you no
longer use, or ask Karmaflow.ai for more.

## Related

- [Voices, Greetings and Pronunciation](/help/voice-agents/voices-and-speech)
- [Multilingual Calls and Language Detection](/help/voice-agents/languages)
- [Calls, Transcripts and Recordings](/help/voice-agents/call-detail)
- [Testing, Drafts and Going Live](/help/voice-agents/testing-and-promotion)