Voice designPreview
Describe a voice in words with voice_description, and Alania-2 creates a new voice for that request. No recording needed.
Writing a description
- Describe age, gender, timbre, pace and energy in a few words.
- Write it in Turkish or English. A description that is not in English is translated to English before it reaches the model.
- At most 200 characters. No parentheses.
- Examples:
A young man with a clear, energetic voice·A calm middle-aged woman with a warm, low voice·tok sesli, yetmişlerinde bir dede(Turkish).
Rules
- When
voice_descriptionis present,voiceis ignored. The OpenAI SDKs always sendvoice, so this is deliberate. - It cannot be combined with
instructionsorreference_audio(conflicting_fields). - A designed voice is not saved on its own; you can save one you like. Every part of a long text is read in the same voice.
- To get the same result again, send a
seed: the same seed, description and text give the same audio. - The default
cfgis 3.0 for design (2.0 otherwise); a higher value holds the voice closer to the description.
Save a designed voice
Save a design you like to your account; after that each request names it by its id in voice, like a saved clone. Send the design request itself (voice_description, seed, input) and a name to POST /v1/voices. The server makes the design's first sentence again with the same seed, has Duyu listen to it, and keeps it as the voice's recording.
POST
https://voice.patientdesk.ai/v1/voicesnamestringrequired- The voice's name, at most 80 characters.
voice_descriptionstringrequired- The design's description, the same as when you rendered it.
seedintegerrequired- The seed that made the voice. The same description gives a different voice with every seed, so a design cannot be saved without one (
seed_required). The Studio picks and keeps a seed for every design. inputstringrequired- The text the design read. What is kept is its first sentence (the part the design made); if that is under 3 seconds, a neutral passage in the same voice is added after it.
cfg · temperature · pronunciations- Send them if you rendered the design with them.
- No real person's voice is in a designed voice, so no
consentis needed. The voice is markedsource: "design"and its description is kept. - Why it is made again from your request: in our measurements, a voice saved from the design you heard stayed clearly closer to it on new sentences, while a passage made from scratch with the same description and seed came out as a different voice.
- A saved design also takes
instructionsand ready styles, and follows a clone's rules: the 50-voice limit, listing withGET /v1/voices, deleting withDELETE /v1/voices/{id}. - If the remade voice reads its description aloud or garbles the text, it is not saved (
design_unusable, 422): render the design again with another seed and save that one.
curl