Skip to main content
POST
Creates a voice from a reference recording and its exact transcript and saves it to your account. Voices are cloned from both together — cloning from the clip alone (x-vector-only) is not supported. The returned id can be passed as voice to Generate speech and the streaming endpoint right away, and it is listed in /v1/models with owned_by: "user". Clones are private. Only API keys on the account that created a clone can use or delete it.
Each account can keep up to 10 cloned voices. Delete one with DELETE /v1/voices/{id} to make room for another. Cloning is English-only.

Authorization

string
required
Bearer sk_… — your secret API key.

Body

string
required
A display name for the voice, up to 40 characters. Must be unique among your voices (case-insensitive).
string
required
The reference clip as base64-encoded WAV, 3–12 seconds long. Use clean speech from a single speaker with no music or background noise. The request body may be up to 16 MB.
string
required
The exact transcript of the reference clip, word for word — fillers (um, uh) and repeated words included, nothing that isn’t spoken. Up to 1000 characters; surrounding whitespace is trimmed. The clip should end where the transcript ends: audio past the last transcribed word can surface as stray words in the generated speech.
Must be true. This confirms you have permission to clone the voice in the clip.
x_vector_only_mode and speaker_embedding are not accepted. Sending either (other than x_vector_only_mode: false) returns 400.

Response

string
The new voice id. Pass this as voice when generating speech.
string
Always voice.
string
The display name, with surrounding whitespace trimmed.
number
Unix timestamp (seconds) the voice was created.
string
Always user.
number
Length of the reference clip in seconds.

Errors