Clone a voice and generate speech with the Percify API

Clone a voice from a short sample with zonos2, or use a preset text to speech voice, through POST /v1/run. Inputs, pricing and example code.

The Percify API clones a voice with the zonos2 model: send the text to speak and a URL of a short recording of the voice, and the generation returns an audio file of that voice saying your text. For a ready-made voice instead of a clone, the same POST /v1/run endpoint runs text to speech models such as speech-02-hd and chatterbox-turbo. The REST API does not store voices, so you send the voice sample, or pick the preset voice, on each call.

Clone a voice with zonos2

POST https://api.percify.io/v3/playground/v1/run

modelId · string · required

zonos2

input.text · string · required

What the cloned voice should say.

input.audio · string · required

Public https URL of a short .mp3 or .wav recording of the voice to clone. Clean speech with little background noise works best.

input.clean_speaker_background · boolean · default: "false"

Remove background noise from the recording before cloning.

cURL

curl -s -X POST https://api.percify.io/v3/playground/v1/run \
  -H "Authorization: Bearer $PERCIFY_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: welcome-line-de" \
  -d '{
    "modelId": "zonos2",
    "input": {
      "text": "Guten Tag und willkommen bei unserem Kurs.",
      "audio": "https://example.com/voice-sample.wav",
      "clean_speaker_background": true
    }
  }'

Wait for the result with GET /v1/generations/{id}?wait=45. The audio file URL is in output.urls[0], with output.type set to audio.

How is voice cloning priced?

zonos2 is priced by the length of the voice sample you pass in audio, not by the length of the text, so a longer sample can cost more. Percify measures the file itself, so call POST /v1/estimate with the same input for the exact credits.

Other voice models

The catalog has more speech models. Each takes its own inputs, listed in input_schema from GET /v1/models/{id}.

Model id	How you choose the voice	Notes from its schema	Credits per run on 16 Sep 2026
chatterbox-turbo	voice preset, or reference_audio to clone	Text up to 500 characters; reference audio must be longer than 5 seconds	5
xtts-v2	speaker: an audio sample to clone	language option	10
speech-02-hd	voice_id system voice	Text up to 10,000 characters; emotion, speed, pitch	5
speech-02-turbo	voice_id system voice	Same inputs as speech-02-hd	5
gemini-tts	voice preset	style_instructions for delivery; priced by text length	Varies
qwen3-tts-flash	voice preset	language_type option	2
seed-speech-tts-2	voice preset	voice_instruction for tone and pace	3

Models without a pricing object in the catalog charge the same credits for every run. Check GET /v1/models for current prices.

Use the audio in a talking avatar

A generated voice track can go straight into lip-sync: pass output.urls[0] as input.audio to infinitetalk-fast or infinitetalk. See Make a talking avatar video.

Tips

• Reuse one sample for a consistent voice across many clips.
• Only clone voices you have permission to use.
• Host samples on a public https URL. If Percify cannot download the zonos2 sample, the run stops with audio URL is not reachable before anything is charged.

Related

Talking avatar pipeline

Turn the voice into a lip-synced video.

Estimate cost

Price a clone before you run it.

Voice cloning in the app

Clone and use a voice without code.

Voice clone troubleshooting

Fixes when a cloned voice sounds wrong.