Voice Cloning API — Check the Integration Contract

Reference-based speech requires the chosen model’s audio and transcript fields. A text-only request does not identify a reference speaker, and voice parameters cannot be copied between models.

Voice Cloning API is a model-family pricing lane on KeyoAPI — a prepaid AI API gateway where one Bearer key and one wallet ($1 = 500,000 quota units) cover chat, speech, vision, OCR and video models.

Documented speech endpoints

Indicative figures for planning. Confirm live Keyo sell rates on interactive Model Square or the static pricing list.

Before using a reference voice

Replace reference URL placeholders with real downloadable files and their exact transcripts. Index uses prompt_audio_url; Cosy uses prompt_wav_url plus prompt_text. Do not invent enrollment or reusable voice-ID fields.

Rates and endpoints

CosyVoice3 uses POST /v1/async/audio/speech with string inputs and required prompt_wav_url plus exact prompt_text. Reference audio must be at most 30 seconds, at least 16 kHz, WAV or MP3. Replace the example URL before submitting; it is not a working reference sample.

IndexTTS-2 requires prompt_audio_url. Its sync request uses input; its async request uses string inputs. Text emotion mode uses use_emo_text and optional emo_text; audio emotion mode instead requires emo_audio_prompt_url. GLM reference examples use prompt_audio_url and matching prompt_text. Step reference fields are optional and WAV-only. Full examples: the API endpoint reference.

Qwen3-TTS uses an inputs array with prompt and language in each item. Built-in speaker names are separate from reference mode; prompt_audio_url and prompt_text belong inside an array item. Spark has separate basic gender/pitch/speed and reference-audio modes. Do not replace these contracts with generic voice or reference_audio fields.

Save id/task_id and query singular GET /v1/task/{id}. On completed, inspect the actual output. Download from url or output.file_url only when present, handling batch results item by item. Preserve an unexpected complete response for support; do not automatically submit again. Retry a temporary task_query_unavailable GET according to Retry-After without creating another task. A submission timeout without an ID has an unknown outcome; do not resubmit automatically.

Style instructions and references do not guarantee a particular emotion, gender, accent or quality. These URL-based requests do not define sample enrollment or reusable voice IDs. Confirm those separate workflows with support. See the TTS API guide for endpoints, the API endpoint reference for model-specific fields and the pricing page for current rates.

# Replace the URL and transcript with your real reference (not a working sample).
curl https://www.keyoapi.xyz/v1/async/audio/speech \
  -H "Authorization: Bearer $KEYO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"CosyVoice3","inputs":"Hello from KeyoAPI TTS.","prompt_wav_url":"https://replace-me.invalid/your-reference.wav","prompt_text":"Replace with the exact transcript of your reference audio."}'
# Save id/task_id, then GET /v1/task/{id}.

Get started in five steps

1. Create an account — new users receive $10 of gift credit for prototyping.
2. Create an API key in the console.
3. Point your OpenAI SDK at https://www.keyoapi.xyz/v1 with the model ID from this page, or call the dedicated REST path from the API handbook.
4. Track spend in the console Wallet — $1 equals 500,000 quota units, and gift credit burns before recharged balance.
5. Set a per-key spending cap before wiring production traffic.

A verified first call is the acceptance test for the setup: the response should carry the payload shape documented in the handbook, the Wallet ledger should show a deduction in quota units (500,000 units = $1), and repeated calls should deduct linearly at the listed rate. Any mismatch usually means the model ID or endpoint is wrong — check both against the pricing list before debugging client code.

FAQ

Also see the model comparison page, the gift-credit models page, the full pricing list, and Text to Speech · CosyVoice3 · Qwen3-TTS.