Voice Remixing

Create a new voice from an existing voice and an editable instruction.

Voice Remixing creates a same-language voice candidate from your own voice or an available official voice. A third-party favorite does not grant permission to remix it. The source voice stays unchanged.

Create a candidate

Send POST /v1/voice-previews/remix with voice_id, prompt, and optional guidance_scale (1–10, default 2). The server freezes the reference audio, transcript, language, prompt and model route. Audio URLs and language overrides are not accepted.

curl -X POST https://api.breeze.blue/v1/voice-previews/remix \
  -H "xi-api-key: $BREEZE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: narrator-older-01" \
  -d '{"voice_id":"voc_example","prompt":"Make the voice older and softer.","guidance_scale":2}'

A 202 response contains generation_job_id and status. Each request creates one candidate. Each successful candidate uses the Voice Design candidate price (100 credits); failed candidates release their reservation. Previewing and saving do not charge again.

Use the same idempotency key to retry an uncertain submission. Identical input returns the original job; changed input returns CONFLICT. If that key is still being submitted, retry the same key after the concurrent request finishes. Use a new key when intentionally generating another candidate.

Poll, audition, save

Poll GET /v1/voice-previews/remix/{generation_job_id}. Once status is ready, generated_voice_id identifies the preview. text is the matching audition script and language_code is inherited from the source.

• Download or audition with GET /v1/voice-previews/{generated_voice_id}/stream.
• Save with POST /v1/voice-previews/{generated_voice_id}/save, supplying voice_name and optional voice_description.
• Use the saved voice_id in Text to Speech with a different script to check the resulting voice.

Saving binds the new audio and its transcript on the server. It creates a private, independent voice and preserves the source task. A candidate must be ready before it can be saved. A polling timeout does not cancel the server task; keep its ID and query it again.

Reference and instruction requirements

The source must have matching audio and a transcript, including a usable preview script of 2–500 characters for comparison. Missing materials block generation before credits are reserved. The original and candidate use the same audition script without generating a separately billed original sample.

Keep the prompt in a language supported by the source model’s instruction contract. Remix does not translate the prompt or change the voice language. Use Voice Localize to create a voice in another language.