Voice cloning

Overview

Voice cloning lets you reproduce a specific speaker from a reference clip. POST /v1/audio/speech takes the clip as ref_audio with its transcript in ref_text, and the model speaks input in that voice.

There are two ways to supply the clip. Inline, the clip travels in the body of every speech request. As a created voice, POST /v1/audio/voices stores the clip once, and the voice's id goes in voice afterwards.

Cloning reproduces a real person, so you need that person's audio and their permission. For a distinctive voice that is not a specific person's, see Voice design.

Clone from a reference clip

curl https://api.sprag.ai/v1/audio/speech \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SPRAG_API_KEY" \
  -d '{
    "model": "chorus",
    "input": "This sentence is spoken in the cloned voice.",
    "ref_audio": "data:audio/wav;base64,UklGRiQAAABXQVZF...",
    "ref_text": "The quick brown fox jumps over the lazy dog.",
    "response_format": "mp3"
  }' \
  --output cloned.mp3

Send ref_audio and ref_text without a voice.

Create a cloned voice

To create a voice from a clip, send the clip as a multipart upload to POST /v1/audio/voices with its transcript, a name, and the model that speaks it.

curl https://api.sprag.ai/v1/audio/voices \
  -H "Authorization: Bearer $SPRAG_API_KEY" \
  -F model=chorus \
  -F [email protected] \
  -F "ref_text=The quick brown fox jumps over the lazy dog." \
  -F name=Priya \
  -F consent_ref=release-2026-0142

The response is a voice with category set to cloned.

{
  "object": "audio.voice",
  "id": "voice_4c7d...",
  "name": "Priya",
  "created_at": 1758067200,
  "description": null,
  "category": "cloned"
}

A name already in use returns 409. A clip that is empty or cannot be measured returns 400. A clip too large to synthesize from returns 413. If your organization already holds as many voices as its plan allows, the request returns HTTP 403 with the error code voice_limit_exceeded.

Speak in a cloned voice

Pass the voice's id as voice on POST /v1/audio/speech, with model set to chorus and no ref_audio. A created voice and inline ref_audio cannot be combined in one request.

curl https://api.sprag.ai/v1/audio/speech \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SPRAG_API_KEY" \
  -d '{
    "model": "chorus",
    "input": "This sentence is spoken in the cloned voice.",
    "voice": "voice_4c7d...",
    "response_format": "mp3"
  }' \
  --output priya.mp3

If the request carries your API key, GET /v1/audio/voices?model=chorus lists the voices your organization created after the presets. DELETE /v1/audio/voices/{voice_id} deletes the voice and its stored clip.

Write the transcript

Set ref_text to what is said in ref_audio, word for word.

A clip you can transcribe exactly is a better reference than a longer one transcribed approximately. When you trim a recording, cut at a boundary you can transcribe with confidence.

If you do not have a transcript, produce one with speech to text and check it before sending it.

Encode the reference clip

ref_audio takes a base64 data: URL, so the clip is inlined into the request body. POST /v1/audio/voices takes the file itself as a multipart upload and needs no encoding.

REF_B64=$(base64 < reference.wav | tr -d '\n')

Common containers such as WAV, MP3, FLAC, OGG, M4A, AAC, AIFF, and WebM are accepted. A clean recording of a single speaker, consistent in tone and level, clones better than a noisy or highly dynamic one. An inline clip travels in every request, so a shorter clip means a smaller request body.

Consent

Treat a reference clip like any other biometric sample. Get explicit permission from the speaker for the use you intend, hold a record of it, and do not reuse a clip for a purpose the speaker did not agree to. The terms of service make that responsibility yours. On a created voice, consent_ref stores a pointer to that record.