Clone a voice for your AI phone agent

An agent can speak in a voice cloned from your own recordings: yours, a colleague’s, or a voice artist’s who has agreed to it. You send the recordings once, get back a voice id, and put that id in an agent’s voice.tts.voice. Four steps:

Make the clone with POST /v1/voices.
Listen to it with GET /v1/voices/{id}/preview.
Put its provider_voice_id in an agent’s voice.
Place a call, or a test call, with that agent.

The console does the same under Voices, at voice.wixzel.com/voices, where you can record yourself in the browser or upload a clip.

Which engines can use a clone

A clone is made at one provider, and only that provider’s engines can speak in it. Choose the provider with provider when you make the clone.

`provider`	Speaks on	Recordings	Languages
`elevenlabs` (default)	`classic`, `deepgram_agent`	1 to 5 files; 30 seconds to a few minutes in total	The languages of the agent’s ElevenLabs model
`sarvam`	`sarvam`	1 file; 10 to 30 seconds is enough	Every language Sarvam supports, including Malayalam, whatever language the recording is in

gemini_live cannot speak in a cloned voice. Gemini Live only offers its own prebuilt voices.

To use the same voice on both kinds of engine, clone it twice, once on each provider. Each clone counts toward the limit and is charged separately.

1. Make the clone

Send the recordings as files in a multipart request:

curl https://api.voice.wixzel.com/v1/voices \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -F name="Meera, front desk" \
  -F provider=elevenlabs \
  -F consent=true \
  -F files=@meera-1.wav \
  -F files=@meera-2.wav

Or, from JSON, as links the API downloads once:

curl https://api.voice.wixzel.com/v1/voices \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Meera, Malayalam",
    "provider": "sarvam",
    "language": "ml-IN",
    "consent": true,
    "audio_urls": ["https://example.com/recordings/meera-ml.wav"]
  }'

Use one or the other in a request, not both. Links must be public https or http URLs. Private and local addresses are refused.

Field	Notes
`name`	Required. Up to 100 characters.
`description`	Optional. Up to 500 characters.
`provider`	`elevenlabs` (default) or `sarvam`. Decides the engines, as above.
`language`	The language spoken in the recording, such as `ml-IN` or `hi-IN`. Required for `sarvam`.
`consent`	Required, and must be `true`. See Consent.
`remove_background_noise`	ElevenLabs only. Cleans up a noisy recording. It can make a clean one sound worse, so leave it off unless you need it.
`files`	Multipart only. The recordings.
`audio_urls`	JSON only. Links to the recordings, up to 5.

The response is the voice:

{
  "id": "6703f1c2a9e4b1d2c3f4a5b6",
  "object": "voice",
  "name": "Meera, front desk",
  "description": null,
  "provider": "elevenlabs",
  "provider_voice_id": "c38kUX8pkfYO2kHyqfFy",
  "engines": ["classic", "deepgram_agent"],
  "language": null,
  "created_at": "2026-10-08T09:12:44.000Z",
  "updated_at": "2026-10-08T09:12:44.000Z"
}

id is the voice’s id in this API, for reading, renaming and deleting it. provider_voice_id is the id an agent uses. engines lists where it can speak. The request waits while the provider processes the audio, which usually takes a few seconds and can take up to a minute.

2. Listen to it

curl "https://api.voice.wixzel.com/v1/voices/$VOICE_ID/preview?language=ml-IN" \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -o preview.mp3

The response is an mp3 of the voice saying a short line. language is optional. Without it, the preview is in the recording’s language, or English for an ElevenLabs clone with no language set. A Sarvam clone can preview in any Sarvam language, which is a quick way to hear it speak a language it was not recorded in. Previews are free. The first one in each language is made and then kept, so playing it again is instant.

3. Give it to an agent

Put provider_voice_id in voice.tts.voice, on an agent whose engine is in the voice’s engines:

curl -X PATCH https://api.voice.wixzel.com/v1/agents/$AGENT_ID \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "voice": {
      "stt": { "model": "deepgram/nova-3", "language": "en-US" },
      "llm": { "model": "openrouter/gpt-4o-mini" },
      "tts": { "model": "elevenlabs/eleven_flash_v2_5", "voice": "c38kUX8pkfYO2kHyqfFy" }
    }
  }'

curl -X PATCH https://api.voice.wixzel.com/v1/agents/$AGENT_ID \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "language": "ml-IN",
    "voice": {
      "stt": { "model": "sarvam/saaras-v3-realtime", "language": "ml-IN" },
      "llm": { "model": "sarvam/sarvam-105b" },
      "tts": { "model": "sarvam/bulbul-v3", "voice": "svc-..." }
    }
  }'

The same works on POST /v1/agents. Your clones are also listed first in GET /v1/engines/{engine}/voices, with "category": "cloned", so a voice picker built on that endpoint shows them without extra work. A clone used on an engine that cannot speak it is refused with unsupported_voice, and the message says which engines it does speak on.

4. Call

Place a call with the agent as usual, or check the voice first with a test call, which rings a number and speaks one phrase:

curl https://api.voice.wixzel.com/v1/agents/$AGENT_ID/test-call \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{ "to": "+919876543210" }'

consent: true confirms that the voice is yours, or that the speaker has given you permission to clone it and to use it on calls. A request without it is refused with 400 invalid_parameter. The confirmation is stored with the time, your account and the IP address the request came from. Do not clone a voice you do not have permission to use. See the Acceptable use policy.

Who can see and use a clone

Only the account that made it.

GET /v1/voices and GET /v1/engines/{engine}/voices show an account its own clones and nobody else’s.
An agent create or update that names another account’s clone is refused with unsupported_voice, the same answer as for a voice that does not exist.
The recordings are passed to the provider to make the clone. Wixzel Voice does not store them.

Cost and limits


Making a clone	$1.00, once, from your credit balance
Previews	Free
Speech in a clone	The same per-character price as the provider’s other voices
Clones per account	3

The $1.00 is charged when the clone is made. It appears in /v1/usage/events as a voice_clone line. If your balance cannot cover it, the request is refused with 402 insufficient_credits before anything is sent to the provider. Deleting a clone does not refund the fee. Speech in an ElevenLabs clone is billed like any ElevenLabs voice on the same model. Speech in a Sarvam clone is billed like Bulbul v3: Sarvam prices cloned speech at ₹30 per 10,000 characters, the same as Bulbul v3. The per-minute prices on Voice engines apply unchanged. An account can hold 3 clones. A fourth is refused with 409 account_limit_reached. Delete one you no longer need, or contact support to raise the limit.

On a self-hosted install the clones are made on your own provider accounts and the

1.00 fee is

Recording tips

The clone sounds like the recording, so the recording matters more than anything else.

Length. For ElevenLabs, 30 seconds to a few minutes in total. For Sarvam, one file of 10 to 30 seconds.
One speaker. No other voices, and no music.
A quiet room. No echo, fans or traffic. A phone held close in a small, furnished room is better than a good microphone in an empty one.
Speak the way the agent will. Talk naturally, at the pace and tone of a phone call. A clone of someone reading in a flat voice will sound flat on calls.
Format. WAV or FLAC is best. mp3, m4a, ogg and webm are accepted. Up to 10 MB per file.

Managing voices

# List your voices
curl https://api.voice.wixzel.com/v1/voices \
  -H "Authorization: Bearer $WIXZEL_API_KEY"

# Read one
curl https://api.voice.wixzel.com/v1/voices/$VOICE_ID \
  -H "Authorization: Bearer $WIXZEL_API_KEY"

# Rename it
curl -X PATCH https://api.voice.wixzel.com/v1/voices/$VOICE_ID \
  -H "Authorization: Bearer $WIXZEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "name": "Meera, reception" }'

# Delete it
curl -X DELETE https://api.voice.wixzel.com/v1/voices/$VOICE_ID \
  -H "Authorization: Bearer $WIXZEL_API_KEY"

PATCH changes name and description only. The recordings and the provider cannot be changed; to change them, make a new clone. DELETE deletes the clone at the provider too, and frees one of your 3 places. It is refused with 409 voice_in_use while an agent uses the voice, and the message names the agents. Give them another voice first. The list is cursor-paginated like every other list. See Pagination.

Scopes

Reading voices and previews needs voices:read. Making, renaming and deleting them needs voices:write.

SDKs and the MCP server

The SDKs have the same operations:

const voice = await client.voices.create({
  name: 'Meera, Malayalam',
  provider: 'sarvam',
  language: 'ml-IN',
  consent: true,
  audio_urls: ['https://example.com/recordings/meera-ml.wav'],
});

client.voices.list, retrieve, update, delete and preview cover the rest. The MCP server has the tools list_voices, get_voice, preview_voice, create_voice_clone, update_voice and delete_voice, so you can make a clone from Claude Code, Codex or Cursor. create_voice_clone takes audio_urls, since an MCP client cannot upload a file.

Errors

Status	`code`	Meaning
`400`	`missing_audio`	No recordings. Send `files` in a multipart request or `audio_urls` in JSON.
`400`	`ambiguous_audio`	Both `files` and `audio_urls` were sent. Send one.
`400`	`unsupported_audio`	A file is not audio. Send mp3, wav, m4a, ogg, webm or flac.
`400`	`too_many_files`	More recordings than the provider takes: 5 for `elevenlabs`, 1 for `sarvam`.
`400`	`missing_language`	`provider` is `sarvam` and `language` is missing.
`400`	`invalid_parameter`	A field is invalid. With `param: "consent"`, `consent` was not `true`.
`400`	`unsafe_audio_url`	An `audio_urls` link points to a private or local address.
`400`	`audio_url_unreachable`	An `audio_urls` link could not be downloaded.
`400`	`voice_clone_rejected`	The provider refused the recording. The message gives its reason. Usually too short, too noisy, or more than one speaker.
`402`	`insufficient_credits`	Your balance cannot cover the $1.00 fee. Top up and try again.
`409`	`account_limit_reached`	The account already has 3 clones. Delete one first.
`409`	`voice_in_use`	On `DELETE`: an agent uses this voice. Give it another voice first.
`413`	`file_too_large`	A recording is larger than 10 MB.
`502`	`provider_unavailable`	The provider failed. Nothing was charged. Try again.
`503`	`voice_capacity`	Cloning is unavailable on the platform for now. Nothing was charged. Try again later.

On agent writes (POST /v1/agents, PATCH /v1/agents/{id}):

Status	`code`	Meaning
`400`	`unsupported_voice`	The voice is another account’s clone, does not exist, or is a clone on a provider this engine cannot speak.

See Errors for the error format.

Getting started

Guides

Carriers

Concepts

API reference

Clone a voice for your AI phone agent

Which engines can use a clone

1. Make the clone

2. Listen to it

3. Give it to an agent

4. Call

Who can see and use a clone

Cost and limits

Recording tips

Managing voices

Scopes

SDKs and the MCP server

Errors

​Which engines can use a clone

​1. Make the clone

​2. Listen to it

​3. Give it to an agent

​4. Call

​Consent

​Who can see and use a clone

​Cost and limits

​Recording tips

​Managing voices

​Scopes

​SDKs and the MCP server

​Errors

Which engines can use a clone

1. Make the clone

2. Listen to it

3. Give it to an agent

4. Call

Consent

Who can see and use a clone

Cost and limits

Recording tips

Managing voices

Scopes

SDKs and the MCP server

Errors