Voice Clone API parameters

Reference for POST /v2/voice_clones/from_upload: Qwen3TTS, Ultra, Bayan, Minimax, and Resemble model options, audio length limits, ref_text, and Ultra async polling.

Models

Set via model_id (body) on POST /v2/voice_clones/from_upload, or use provider (body) to select Minimax or Resemble.

Model	Provider	Languages	Audio length	Max file	Sync/Async	ref_text	Best for
Qwen3TTS	Telnyx (default)	10 languages	3–15s (auto-trimmed to 10s)	20 MB	Sync (201)	Optional	Short, clean samples
Ultra	Telnyx	42 languages	Up to 60s (auto-trimmed)	20 MB	Async (202)	Optional	Higher quality, more natural
Bayan	Telnyx	ar (default), en	~10s (auto-trimmed to 10s)	20 MB	Sync (201)	Required	Arabic and English voices
speech-2.8-turbo	Minimax	40+ languages	10s–5 min	20 MB	Sync (201)	Optional	Longer recordings, more vocal range
Pro	Resemble	23 languages (language required)	~10s (first 10s used)	20 MB	Sync (201)	Not used	Multilingual Resemble voices
Turbo	Resemble	English (en, en-US, default en-US)	~10s (first 10s used)	20 MB	Sync (201)	Not used	English voices
Urdu	Resemble	ur (default)	~10s (first 10s used)	20 MB	Sync (201)	Not used	Urdu voices

Audio Requirements

Body parameter audio_file (multipart) on POST /v2/voice_clones/from_upload.

	Qwen3TTS	Ultra	Bayan	Minimax	Resemble
Audio length	3–15s (5–10s optimal)	Up to 60s	~10s (first 10s used)	10s–5 min	~10s (first 10s used)
Max file size	20 MB	20 MB	20 MB	20 MB	20 MB
Formats	WAV, MP3, FLAC, OGG, M4A	Same	Same	Same	Same

• Qwen3TTS: aim for 5–10 seconds. Longer isn’t better — auto-trims to 10s.
• Ultra: up to 60 seconds. Longer, varied speech gives the model more of the voice to learn from.
• Bayan: about 10 seconds of clean, single-speaker speech. The audio is converted to 24 kHz mono WAV and trimmed to the first 10 seconds.
• Minimax: longer is better. 1–2 minutes of varied speech gives more vocal range.
• Resemble: about 10 seconds of clean, single-speaker speech. The audio is converted to WAV and trimmed to the first 10 seconds.

The ref_text Parameter

Body parameter on POST /v2/voice_clones/from_upload. Required for Bayan, ignored for Resemble, optional for every other model.

A transcript of what’s being said in the audio. Improves clone quality by giving the model a text reference to align against.

Bayan requires an exact transcript. A missing, empty, or whitespace-only ref_text returns 422 (code 10027, source.pointer /ref_text). An inaccurate transcript can make synthesis with the clone truncated or garbled. Record a known script of about 10 seconds and pass the script verbatim as ref_text.

Bayan Language

When model_id is "Bayan", language accepts only ar (default) or en. Other values return 422 with source.pointer /language.

Resemble Models

Set provider to resemble and model_id to Pro, Turbo, or Urdu.

• Pro: language is required: ar, da, de, el, en-US, es, fi, fr-fr, he, hi, it, ja, ko, ms, nl, no, pl, pt, ru, sv, sw, tr, zh.
• Turbo: English only. language accepts en or en-US (default en-US).
• Urdu: language accepts ur (default).

Other language values return 422 with source.pointer /language.

A Resemble clone works only with the model it was created for. To use the same speaker on another model, create a separate clone with that model_id.

Ultra Async Flow

When model_id is "Ultra", the API returns 202 Accepted instead of 201:

POST /v2/voice_clones/from_upload → 202 { "data": { "status": "pending" } }

Poll until ready:

GET /v2/voice_clones/{id} → 200 { "data": { "status": "active" } }

See Responses for status values and voice ID format. See Errors for Minimax error codes.